Misplaced Pages

CUDA: Difference between revisions

Article snapshot taken from Wikipedia with creative commons attribution-sharealike license. Give it a read and then ask your questions in the chat. We can research this topic together.
Browse history interactively← Previous editContent deleted Content addedVisualWikitext
Revision as of 03:09, 16 May 2015 editPateljay43 (talk | contribs)29 edits Dont change on any of your assumption. There is nothing like 'compute unified device architecture' in that reference because NVidia never gave that name. That name is big wrong assumption carried on.← Previous edit Latest revision as of 17:23, 11 January 2025 edit undoMaxeto0910 (talk | contribs)Extended confirmed users95,300 edits →GPUs supportedTag: Visual edit 
(966 intermediate revisions by more than 100 users not shown)
Line 1: Line 1:
{{Short description|Parallel computing platform and programming model}}
{{multiple issues|
{{how-to|date=February 2024}}
{{Pro and con list|date=February 2024}}
}}
{{Infobox software {{Infobox software
| name = CUDA | name = CUDA
| screenshot = | screenshot = Nvidia CUDA Logo.jpg
| developer = ]
| caption = A parallel computing platform and programming model
| released = {{Start date and age|2007|02|15}}
| developer = ]
| latest_release_version = 12.6
| released = {{start date and age|2007|06|23}}
| latest_release_date = {{Start date and age|2024|08}}
| latest_release_version = 7.0
| operating_system = ], ]
| latest_release_date = {{Start date and age|2015|03|17}}
| platform = ]
| operating_system = ] and ],<br />], ]
| genre = ]
| platform = ]
| license = ]
| genre = ]
| website = {{URL|https://developer.nvidia.com/cuda-zone}}
| license = ]
| website = {{URL|www.nvidia.com/object/cuda_home_new.html}}
}} }}
In ], '''CUDA''' is a proprietary<ref name=":0">{{Cite web |last=Shah |first=Agam |title=Nvidia not totally against third parties making CUDA chips |url=https://www.theregister.com/2021/11/10/nvidia_cuda_silicon/ |access-date=2024-04-25 |website=www.theregister.com |language=en}}</ref> ] platform and ] (API) that allows software to use certain types of ] (GPUs) for accelerated general-purpose processing, an approach called general-purpose computing on GPUs. CUDA was created by ] in 2006.<ref>{{cite web |date=18 July 2017 |title=Nvidia CUDA Home Page |url=https://developer.nvidia.com/cuda-zone}}</ref> When it was first introduced, the name was an acronym for Compute Unified Device Architecture,<ref name="CUDA intro - AnandTech">{{cite web |last1=Shimpi |first1=Anand Lal |last2=Wilson |first2=Derek |date=November 8, 2006 |title=Nvidia's GeForce 8800 (G80): GPUs Re-architected for DirectX&nbsp;10 |url=https://www.anandtech.com/show/2116/8 |access-date=May 16, 2015 |publisher=AnandTech}}</ref> but Nvidia later ] the common use of the acronym and now rarely expands it.<ref>{{Cite web |title=Introduction — nsight-visual-studio-edition 12.6 documentation |url=https://docs.nvidia.com/nsight-visual-studio-edition/introduction/index.html#cuda-debugger |access-date=2024-10-10 |website=docs.nvidia.com}}</ref>
'''CUDA''' (after the ]<ref></ref>), is a ] platform and programming model created by ] and implemented by the ]s (GPUs) that they produce.<ref></ref> CUDA gives developers direct access to the virtual ] and memory of the parallel computational elements in CUDA GPUs. CUDA is just name not an acronym <ref>http://docs.nvidia.com/cuda/pdf/CUDA_Runtime_API.pdf</ref> .


CUDA is a software layer that gives direct access to the GPU's virtual ] and parallel computational elements for the execution of ]s.<ref name="CUDA intro - TomsHardware">{{cite web |url=https://www.tomshardware.com/reviews/nvidia-cuda-gpu,1954.html |title=Nvidia's CUDA: The End of the CPU? |last=Abi-Chahla |first=Fedy |date=June 18, 2008 |publisher=Tom's Hardware |access-date=May 17, 2015}}</ref> In addition to ] and runtime kernels, the CUDA platform includes compilers, libraries and developer tools to help programmers accelerate their applications.
Using CUDA, the GPUs can be used for general purpose processing (i.e., not exclusively graphics); this approach is known as ]. Unlike CPUs, however, GPUs have a parallel throughput architecture that emphasizes executing many concurrent threads slowly, rather than executing a single thread very quickly.


CUDA is designed to work with programming languages such as ], ], ] and ]. This accessibility makes it easier for specialists in parallel programming to use GPU resources, in contrast to prior APIs like ] and ], which require advanced skills in graphics programming.<ref>{{Cite news |url=https://www.videomaker.com/article/c15/19313-cuda-vs-opencl-vs-opengl |title=CUDA vs. OpenCL vs. OpenGL |last=Zunitch |first=Peter |date=2018-01-24 |work=Videomaker |access-date=2018-09-16 |language=en-US}}</ref> CUDA-powered GPUs also support programming frameworks such as ], ] and ].<ref>{{Cite web |url=https://developer.nvidia.com/opencl |title=OpenCL |date=2013-04-24 |website=NVIDIA Developer |language=en |access-date=2019-11-04}}</ref><ref name="CUDA intro - TomsHardware" />
The CUDA platform is accessible to software developers through CUDA-accelerated libraries, ] (such as ]), and extensions to industry-standard programming languages, including ], ] and ]. C/C++ programmers use 'CUDA C/C++', compiled with "nvcc", NVIDIA's ]-based C/C++ compiler.<ref></ref> Fortran programmers can use 'CUDA Fortran', compiled with the PGI CUDA Fortran compiler from ].


==Background==
In addition to libraries, compiler directives, CUDA C/C++ and CUDA Fortran, the CUDA platform supports other computational interfaces, including the ]'s ],<ref>{{YouTube|r1sN1ELJfNo|First OpenCL demo on a GPU}}</ref> Microsoft's ], and ].<ref>{{YouTube|K1I4kts5mqc|DirectCompute Ocean Demo Running on Nvidia CUDA-enabled GPU}}</ref> Third party wrappers are also available for ], ], ], ], ], ], ], ], ], ], and native support in ].
{{More information|Graphics processing unit}}


The graphics processing unit (GPU), as a specialized computer processor, addresses the demands of ] high-resolution ] compute-intensive tasks. By 2012, GPUs had evolved into highly parallel ] systems allowing efficient manipulation of large blocks of data. This design is more effective than general-purpose ] (CPUs) for ]s in situations where processing large blocks of data is done in parallel, such as:
In the ] industry, GPUs are used not only for graphics rendering but also in ] (physical effects such as debris, smoke, fire, fluids); examples include ] and ]. CUDA has also been used to accelerate non-graphical applications in ], ] and other fields by an ] or more.<ref name=Ioannidis08>{{cite journal
* ]s
| author = Giorgos Vasiliadis, Spiros Antonatos, Michalis Polychronakis, Evangelos P. Markatos and Sotiris Ioannidis
* ]
| title = Gnort: High Performance Network Intrusion Detection Using Graphics Processors
* ] simulations
| journal = Proceedings of the 11th International Symposium on Recent Advances in Intrusion Detection (RAID)
* ]s
| date=September 2008
Ian Buck, while at Stanford in 2000, created an 8K gaming rig using 32 GeForce cards, then obtained a DARPA grant to perform ]. He then joined Nvidia, where since 2004 he has been overseeing CUDA development. In pushing for CUDA, ] aimed for the Nvidia GPUs to become a general hardware for scientific computing. CUDA was released in 2007. Around 2015, the focus of CUDA changed to neural networks.<ref>{{Cite magazine |last=Witt |first=Stephen |date=2023-11-27 |title=How Jensen Huang's Nvidia Is Powering the A.I. Revolution |language=en-US |magazine=The New Yorker |url=https://www.newyorker.com/magazine/2023/12/04/how-jensen-huangs-nvidia-is-powering-the-ai-revolution |access-date=2023-12-10 |issn=0028-792X}}</ref>
| url = http://www.ics.forth.gr/dcs/Activities/papers/gnort.raid08.pdf
| format=PDF}}</ref><ref>{{cite journal
| author = Schatz, M.C., Trapnell, C., Delcher, A.L., Varshney, A.
| year = 2007
| url = http://www.biomedcentral.com/1471-2105/8/474
| title = High-throughput sequence alignment using Graphics Processing Units
| journal = BMC Bioinformatics
| volume = 8:474
| doi = 10.1186/1471-2105-8-474
| pages = 474
| pmid = 18070356
| pmc = 2222658
}}</ref><ref name=Manavski2008>{{cite journal
| first = Svetlin A.
| last = Manavski
| author2=Giorgio Valle
| title = CUDA compatible GPU cards as efficient hardware accelerators for Smith-Waterman sequence alignment
| url = http://www.biomedcentral.com/1471-2105/9/S2/S10
| journal = BMC Bioinformatics
| volume = 9 | year = 2008
| doi = 10.1186/1471-2105-9-S2-S10
| pages = S10
| pmid = 18387198
| pmc = 2323659
}}</ref><ref>Pyrit – Google Code https://code.google.com/p/pyrit/</ref><ref>, BOINC official site (December 18, 2008)</ref>


== Ontology ==
CUDA provides both a low level ] and a higher level API. The initial CUDA ] was made public on 15 February 2007, for ] and ]. ] support was later added in version 2.0,<ref></ref> which supersedes the beta released February 14, 2008.<ref>- (Posted on Feb 14, 2008)</ref> CUDA works with all Nvidia GPUs from the G8x series onwards, including ], ] and the ] line. CUDA is compatible with most standard operating systems. Nvidia states that programs developed for the G8x series will also work without modification on all future Nvidia video cards, due to binary compatibility.{{Citation needed|date=January 2014}}
The following table offers a non-exact description for the ] of CUDA framework.


{| class="wikitable"
]
|+ The ontology of CUDA framework
! memory<br>(hardware)
! memory (code, or ])
! computation<br>(hardware)
! computation<br>(code syntax)
! computation<br>(code semantics)
|-
| ]
| non-CUDA variables
| host
| program
| one ]
|-
| ],<br>GPU L2 cache
| global, const, texture
| device
| grid
| simultaneous call of the same ] on many processors
|-
| GPU L1 cache
| local, shared
| SM ("streaming multiprocessor")
| block
| individual subroutine call
|-
|
|
| warp = 32 threads
|
| ]s
|-
| GPU L0 cache,<br>register
|
| thread (aka. "SP", "streaming processor", "cuda core", but these names are now deprecated)
|
| analogous to individual scalar ops within a vector op
|}


==Programming abilities==
==Background==
] |3=GPU's CUDA cores execute the kernel in parallel |4=Copy the resulting data from GPU memory to main memory}}]]
{{See also|GPU}}


The CUDA platform is accessible to software developers through CUDA-accelerated libraries, ] such as ], and extensions to industry-standard programming languages including ], ], ] and ]. C/C++ programmers can use 'CUDA C/C++', compiled to ] with ], Nvidia's ]-based C/C++ compiler, or by clang itself.<ref>{{cite web|url=https://developer.nvidia.com/cuda-llvm-compiler|title=CUDA LLVM Compiler|date=7 May 2012}}</ref> Fortran programmers can use 'CUDA Fortran', compiled with the PGI CUDA Fortran compiler from ].{{Update inline|reason=PGI Compilers & Tools have evolved into the NVIDIA HPC SDK. The current Fortran compiler is called nvfortran.|date=December 2022}} Python programmers can use the cuNumeric library to accelerate applications on Nvidia GPUs.
The GPU, as a specialized ], addresses the demands of ] high-resolution ] compute-intensive tasks. {{As of | 2012}}, GPUs have evolved into highly parallel ] systems allowing very efficient manipulation of large blocks of data. This design is more effective than general-purpose ] for ]s where processing of large blocks of data is done in parallel, such as:


In addition to libraries, compiler directives, CUDA C/C++ and CUDA Fortran, the CUDA platform supports other computational interfaces, including the ]'s ],<ref>{{YouTube|r1sN1ELJfNo|First OpenCL demo on a GPU}}</ref> Microsoft's ], ] Compute Shader and ].<ref>{{YouTube|K1I4kts5mqc|DirectCompute Ocean Demo Running on Nvidia CUDA-enabled GPU}}</ref> Third party wrappers are also available for ], ], Fortran, ], ], ], ], ], ], ], ], ], and native support in ].
* ]

* fast ]s of large ]
In the ] industry, GPUs are used for graphics rendering, and for ] (physical effects such as debris, smoke, fire, fluids); examples include ] and ]. CUDA has also been used to accelerate non-graphical applications in ], ] and other fields by an ] or more.<ref name=Ioannidis08>{{cite book|last1=Vasiliadis |first1=Giorgos |last2=Antonatos |first2=Spiros |last3=Polychronakis |first3=Michalis |last4=Markatos |first4=Evangelos P. |last5=Ioannidis |first5=Sotiris |title=Recent Advances in Intrusion Detection |chapter=Gnort: High Performance Network Intrusion Detection Using Graphics Processors |series=Lecture Notes in Computer Science |date=September 2008 |volume=5230 |pages=116–134 |doi=10.1007/978-3-540-87403-4_7 |isbn=978-3-540-87402-7 |chapter-url= http://www.ics.forth.gr/dcs/Activities/papers/gnort.raid08.pdf }}</ref><ref>{{cite journal |last1=Schatz |first1=Michael C. |last2=Trapnell |first2=Cole |last3=Delcher |first3=Arthur L. |last4=Varshney |first4=Amitabh |year= 2007 |title= High-throughput sequence alignment using Graphics Processing Units |journal= BMC Bioinformatics |volume= 8|doi= 10.1186/1471-2105-8-474 |pages= 474 |pmid= 18070356 |pmc= 2222658 |doi-access=free }}</ref><ref name=Manavski2008>{{cite journal|last1= Manavski |first1= Svetlin A. |last2=Giorgio |first2=Valle |title= CUDA compatible GPU cards as efficient hardware accelerators for Smith-Waterman sequence alignment |journal= BMC Bioinformatics |volume= 10 |year= 2008 |issue= Suppl 2 |doi= 10.1186/1471-2105-9-S2-S10 |pages= S10 |pmid= 18387198 |pmc= 2323659 |doi-access= free }}</ref><ref>{{cite web|url=https://code.google.com/p/pyrit/|title=Pyrit – Google Code}}</ref><ref>{{cite web|url=http://boinc.berkeley.edu/cuda.php|title=Use your Nvidia GPU for scientific computing|archive-url=https://web.archive.org/web/20081228022142/http://boinc.berkeley.edu/cuda.php|archive-date=2008-12-28|url-status=dead|access-date=2017-08-08|publisher=BOINC|date=2008-12-18}}</ref>
* two-dimensional ]

* ] simulations
CUDA provides both a low level ] (CUDA '''Driver''' API, non single-source) and a higher level API (CUDA '''Runtime''' API, single-source). The initial CUDA ] was made public on 15 February 2007, for ] and ]. ] support was later added in version 2.0,<ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/sdk/website/doc/CUDA_SDK_release_notes_macosx.txt|title=Nvidia CUDA Software Development Kit (CUDA SDK) – Release Notes Version 2.0 for MAC OS X|url-status=dead|archive-url=https://web.archive.org/web/20090106020401/http://developer.download.nvidia.com/compute/cuda/sdk/website/doc/CUDA_SDK_release_notes_macosx.txt|archive-date=2009-01-06}}</ref> which supersedes the beta released February 14, 2008.<ref>{{cite web|url=http://news.developer.nvidia.com/2008/02/cuda-11---now-o.html|title=CUDA 1.1 – Now on Mac OS X|date=February 14, 2008|url-status=dead|archive-url=https://web.archive.org/web/20081122105633/http://news.developer.nvidia.com/2008/02/cuda-11---now-o.html|archive-date=November 22, 2008}}</ref> CUDA works with all Nvidia GPUs from the G8x series onwards, including ], ] and the ] line. CUDA is compatible with most standard operating systems.

CUDA 8.0 comes with the following libraries (for compilation & runtime, in alphabetical order):

* cuBLAS – CUDA Basic Linear Algebra Subroutines library
* CUDART – CUDA Runtime library
* cuFFT – CUDA Fast Fourier Transform library
* cuRAND – CUDA Random Number Generation library
* cuSOLVER – CUDA based collection of dense and sparse direct solvers
* cuSPARSE – CUDA Sparse Matrix library
* NPP – NVIDIA Performance Primitives library
* nvGRAPH – NVIDIA Graph Analytics library
* NVML – NVIDIA Management Library
* NVRTC – NVIDIA Runtime Compilation library for CUDA C++
CUDA 8.0 comes with these other software components:

* nView – NVIDIA nView Desktop Management Software
* NVWMI – NVIDIA Enterprise Management Toolkit
* GameWorks ] – is a multi-platform game physics engine

CUDA 9.0–9.2 comes with these other components:

* CUTLASS 1.0 – custom linear algebra algorithms,
* NVIDIA Video Decoder was deprecated in CUDA 9.2; it is now available in NVIDIA Video Codec SDK

CUDA 10 comes with these other components:

* nvJPEG – Hybrid (CPU and GPU) JPEG processing

CUDA 11.0–11.8 comes with these other components:<ref>{{Cite web|url=https://developer.nvidia.com/blog/cuda-11-features-revealed/|title=CUDA 11 Features Revealed|date=14 May 2020}}</ref><ref>{{Cite web|url=https://developer.nvidia.com/blog/cuda-11-1-introduces-support-rtx-30-series/|title = CUDA Toolkit 11.1 Introduces Support for GeForce RTX 30 Series and Quadro RTX Series GPUs|date = 23 September 2020}}</ref><ref>{{Cite web|url=https://developer.nvidia.com/blog/enhancing-memory-allocation-with-new-cuda-11-2-features/|title=Enhancing Memory Allocation with New NVIDIA CUDA 11.2 Features|date=16 December 2020}}</ref><ref>{{Cite web|url=https://developer.nvidia.com/blog/exploring-the-new-features-of-cuda-11-3/|title = Exploring the New Features of CUDA 11.3|date = 16 April 2021}}</ref>

* CUB is new one of more supported C++ libraries
* MIG multi instance GPU support
* nvJPEG2000 – ] encoder and decoder


==Advantages== ==Advantages==
CUDA has several advantages over traditional general-purpose computation on GPUs (GPGPU) using graphics APIs: CUDA has several advantages over traditional general-purpose computation on GPUs (GPGPU) using graphics APIs:


* Scattered reads{{snd}} code can read from arbitrary addresses in memory * Scattered reads{{snd}}code can read from arbitrary addresses in memory.
* Unified virtual memory (CUDA&nbsp;4.0 and above) * Unified virtual memory (CUDA&nbsp;4.0 and above)
* Unified memory (CUDA&nbsp;6.0 and above) * Unified memory (CUDA&nbsp;6.0 and above)
* ]{{snd}} CUDA exposes a fast ] region that can be shared amongst threads. This can be used as a user-managed cache, enabling higher bandwidth than is possible using texture lookups.<ref>{{cite conference|doi=10.1145/1375527.1375572|title=Efficient computation of sum-products on GPUs through software-managed cache|conference=Proceedings of the 22nd annual international conference on Supercomputing - ICS '08|year=2008|last1=Silberstein|first1=Mark|last2=Schuster|first2=Assaf|last3=Geiger|first3=Dan|last4=Patney|first4=Anjul|last5=Owens|first5=John D.|isbn=978-1-60558-158-3|pages=309–318}}</ref> * ]{{snd}}CUDA exposes a fast shared memory region that can be shared among threads. This can be used as a user-managed cache, enabling higher bandwidth than is possible using texture lookups.<ref>{{cite conference|doi=10.1145/1375527.1375572|chapter=Efficient computation of sum-products on GPUs through software-managed cache|conference=Proceedings of the 22nd annual international conference on Supercomputing – ICS '08|year=2008|last1=Silberstein|first1=Mark|last2=Schuster|first2=Assaf|author2-link= Assaf Schuster
|last3=Geiger|first3=Dan|last4=Patney|first4=Anjul|last5=Owens|first5=John D.|title=Proceedings of the 22nd annual international conference on Supercomputing – ICS '08|isbn=978-1-60558-158-3|pages=309–318|chapter-url=https://escholarship.org/content/qt8js4v3f7/qt8js4v3f7.pdf?t=ptt3te|url=https://escholarship.org/content/qt8js4v3f7/qt8js4v3f7.pdf?t=ptt3te}}</ref>
* Faster downloads and readbacks to and from the GPU * Faster downloads and readbacks to and from the GPU
* Full support for integer and bitwise operations, including integer texture lookups * Full support for integer and bitwise operations, including integer texture lookups


==Limitations== ==Limitations==
* CUDA does not support the full C standard, as it runs host code through a C++ compiler, which makes some valid C (but invalid C++) code fail to compile.<ref></ref><ref></ref> * Whether for the host computer or the GPU device, all CUDA source code is now processed according to C++ syntax rules.<ref name="CUDA_Prog_v8">{{cite web<!--|title=CUDA Toolkit Documentation-->|url=http://docs.nvidia.com/cuda/pdf/CUDA_C_Programming_Guide.pdf|website=nVidia Developer Zone |title=CUDA C Programming Guide v8.0|access-date=22 March 2017|page=19|date=January 2017}}</ref> This was not always the case. Earlier versions of CUDA were based on C syntax rules.<ref>{{cite web|url=https://devtalk.nvidia.com/default/topic/508479/cuda-programming-and-performance/nvcc-forces-c-compilation-of-cu-files/#entry1340190|title=NVCC forces c++ compilation of .cu files|date=29 November 2011}}</ref> As with the more general case of compiling C code with a C++ compiler, it is therefore possible that old C-style CUDA source code will either fail to compile or will not behave as originally intended.
* Interoperability with rendering languages such as OpenGL is one-way, with OpenGL having access to registered CUDA memory but CUDA not having access to OpenGL memory. * Interoperability with rendering languages such as OpenGL is one-way, with OpenGL having access to registered CUDA memory but CUDA not having access to OpenGL memory.
* Copying between host and device memory may incur a performance hit due to system bus bandwidth and latency (this can be partly alleviated with asynchronous memory transfers, handled by the GPU's DMA engine) * Copying between host and device memory may incur a performance hit due to system bus bandwidth and latency (this can be partly alleviated with asynchronous memory transfers, handled by the GPU's DMA engine).
* Threads should be running in groups of at least 32 for best performance, with total number of threads numbering in the thousands. Branches in the program code do not affect performance significantly, provided that each of 32 threads takes the same execution path; the ] execution model becomes a significant limitation for any inherently divergent task (e.g. traversing a ] data structure during ]). * Threads should be running in groups of at least 32 for best performance, with total number of threads numbering in the thousands. Branches in the program code do not affect performance significantly, provided that each of 32 threads takes the same execution path; the ] execution model becomes a significant limitation for any inherently divergent task (e.g. traversing a ] data structure during ]).
* No emulation or fallback functionality is available for modern revisions.
* Unlike ], CUDA-enabled GPUs are only available from Nvidia<ref name="CUDA_products" >{{cite web |url=http://www.nvidia.com/object/cuda_learn_products.html |title=CUDA-Enabled Products |work=CUDA Zone |publisher=Nvidia Corporation |accessdate=2008-11-03}}</ref>
* Valid C++ may sometimes be flagged and prevent compilation due to the way the compiler approaches optimization for target GPU device limitations.{{citation needed|date=May 2016}}
* No emulator or fallback functionality is available for modern revisions
* C++ ] (RTTI) and C++-style exception handling are only supported in host code, not in device code.
* Valid C/C++ may sometimes be flagged and prevent compilation due to optimization techniques the compiler is required to employ to use limited resources.
* In ] on first generation CUDA compute capability 1.x devices, ]s are unsupported and are instead flushed to zero, and the precision of both the division and square root operations are slightly lower than IEEE 754-compliant single precision math. Devices that support compute capability 2.0 and above support denormal numbers, and the division and square root operations are IEEE 754 compliant by default. However, users can obtain the prior faster gaming-grade math of compute capability 1.x devices if desired by setting compiler flags to disable accurate divisions and accurate square roots, and enable flushing denormal numbers to zero.<ref>{{Cite web|url=https://developer.nvidia.com/sites/default/files/akamai/cuda/files/NVIDIA-CUDA-Floating-Point.pdf |first1=Nathan |last1=Whitehead |first2=Alex |last2=Fit-Florea |title=Precision & Performance: Floating Point and IEEE 754 Compliance for Nvidia GPUs |access-date=November 18, 2014 |publisher=]}}</ref>
* A single process must run spread across multiple disjoint memory spaces, unlike other C language runtime environments.
* Unlike ], CUDA-enabled GPUs are only available from Nvidia as it is proprietary.<ref name="CUDA_products" >{{cite web |url=https://www.nvidia.com/object/cuda_learn_products.html |title=CUDA-Enabled Products |work=CUDA Zone |publisher=Nvidia Corporation |access-date=2008-11-03}}</ref><ref name=":0" /> Attempts to implement CUDA on other GPUs include:
* C++ Run-Time Type Information (]) is not supported in CUDA code, due to lack of support in the underlying hardware.
** Project Coriander: Converts CUDA C++11 source to OpenCL 1.2 C. A fork of CUDA-on-CL intended to run ].<ref>{{cite web|url=http://www.phoronix.com/scan.php?page=news_item&px=CUDA-On-CL-Coriander|title=Coriander Project: Compile CUDA Codes To OpenCL, Run Everywhere|publisher=Phoronix}}</ref><ref>{{cite web|url=http://www.iwocl.org/wp-content/uploads/iwocl2017-hugh-perkins-cuda-cl.pdf|title=cuda-on-cl|last=Perkins|first=Hugh|publisher=IWOCL|date=2017|access-date=August 8, 2017}}</ref><ref>{{cite web|url=https://github.com/hughperkins/coriander|title=hughperkins/coriander: Build NVIDIA® CUDA™ code for OpenCL™ 1.2 devices|publisher=GitHub|date=May 6, 2019}}</ref>
* Exception handling is not supported in CUDA code due to performance overhead that would be incurred with many thousands of parallel threads running.
** CU2CL: Convert CUDA 3.2 C++ to OpenCL C.<ref>{{cite web |title=CU2CL Documentation |url=http://chrec.cs.vt.edu/cu2cl/documentation.php |website=chrec.cs.vt.edu}}</ref>
* CUDA (with compute capability 2.x) allows a subset of C++ class functionality, for example member functions may not be virtual (this restriction will be removed in some future release). ''''
** ] HIP: A thin abstraction layer on top of CUDA and ] intended for AMD and Nvidia GPUs. Has a conversion tool for importing CUDA C++ source. Supports CUDA 4.0 plus C++11 and float16.
* In ] on first generation CUDA compute capability 1.x devices, ]s are not supported and are instead flushed to zero, and the precisions of the division and square root operations are slightly lower than IEEE 754-compliant single precision math. Devices that support compute capability 2.0 and above support denormal numbers, and the division and square root operations are IEEE 754 compliant by default. However, users can obtain the previous faster gaming-grade math of compute capability 1.x devices if desired by setting compiler flags to disable accurate divisions, disable accurate square roots, and enable flushing denormal numbers to zero.<ref>{{Cite web|url=https://developer.nvidia.com/sites/default/files/akamai/cuda/files/NVIDIA-CUDA-Floating-Point.pdf |first1=Nathan |last1=Whitehead |first2=Alex |last2=Fit-Florea |title=Precision & Performance: Floating Point and IEEE 754 Compliance for NVIDIA GPUs |accessdate=November 18, 2014 |publisher=]}}</ref>
** ZLUDA is a drop-in replacement for CUDA on AMD GPUs and formerly Intel GPUs with near-native performance.<ref>{{cite web | title= GitHub – vosen/ZLUDA|website=] |url=https://github.com/vosen/ZLUDA }}</ref> The developer, Andrzej Janik, was separately contracted by both Intel and AMD to develop the software in 2021 and 2022, respectively. However, neither company decided to release it officially due to the lack of a business use case. AMD's contract included a clause that allowed Janik to release his code for AMD independently, allowing him to release the new version that only supports AMD GPUs.<ref>{{Citation |last=Larabel |first=Michael |title=AMD Quietly Funded A Drop-In CUDA Implementation Built On ROCm: It's Now Open-Source |date=2024-02-12 |work=] |url=https://www.phoronix.com/review/radeon-cuda-zluda |access-date=2024-02-12 |language=en}}</ref>
** chipStar can compile and run CUDA/HIP programs on advanced OpenCL 3.0 or Level Zero platforms.<ref>{{cite web | title= GitHub – chip-spv/chipStar|website=] |url=https://github.com/chip-spv/chipStar }}</ref>


==Example==
== Supported GPUs ==
This example code in ] loads a texture from an image into an array on the GPU:
Compute capability table (version of CUDA supported) by GPU and card. Also available directly from :

<syntaxhighlight lang="cuda">
texture<float, 2, cudaReadModeElementType> tex;

void foo()
{
cudaArray* cu_array;

// Allocate array
cudaChannelFormatDesc description = cudaCreateChannelDesc<float>();
cudaMallocArray(&cu_array, &description, width, height);

// Copy image data to array
cudaMemcpyToArray(cu_array, image, width*height*sizeof(float), cudaMemcpyHostToDevice);

// Set texture parameters (default)
tex.addressMode = cudaAddressModeClamp;
tex.addressMode = cudaAddressModeClamp;
tex.filterMode = cudaFilterModePoint;
tex.normalized = false; // do not normalize coordinates

// Bind the array to the texture
cudaBindTextureToArray(tex, cu_array);

// Run kernel
dim3 blockDim(16, 16, 1);
dim3 gridDim((width + blockDim.x - 1)/ blockDim.x, (height + blockDim.y - 1) / blockDim.y, 1);
kernel<<< gridDim, blockDim, 0 >>>(d_data, height, width);

// Unbind the array from the texture
cudaUnbindTexture(tex);
} //end foo()

__global__ void kernel(float* odata, int height, int width)
{
unsigned int x = blockIdx.x*blockDim.x + threadIdx.x;
unsigned int y = blockIdx.y*blockDim.y + threadIdx.y;
if (x < width && y < height) {
float c = tex2D(tex, x, y);
odata = c;
}
}
</syntaxhighlight>

Below is an example given in ] that computes the product of two arrays on the GPU. The unofficial Python language bindings can be obtained from ''PyCUDA''.<ref>{{cite web|url=http://mathema.tician.de/software/pycuda|title=PyCUDA}}</ref>
<syntaxhighlight lang="numpy">
import pycuda.compiler as comp
import pycuda.driver as drv
import numpy
import pycuda.autoinit

mod = comp.SourceModule(
"""
__global__ void multiply_them(float *dest, float *a, float *b)
{
const int i = threadIdx.x;
dest = a * b;
}
"""
)

multiply_them = mod.get_function("multiply_them")

a = numpy.random.randn(400).astype(numpy.float32)
b = numpy.random.randn(400).astype(numpy.float32)

dest = numpy.zeros_like(a)
multiply_them(drv.Out(dest), drv.In(a), drv.In(b), block=(400, 1, 1))

print(dest - a * b)
</syntaxhighlight>

Additional Python bindings to simplify matrix multiplication operations can be found in the program ''pycublas''.<ref>{{cite web|url=http://kered.org/blog/2009-04-13/easy-python-numpy-cuda-cublas/|title=pycublas|archive-url=https://web.archive.org/web/20090420124748/http://kered.org/blog/2009-04-13/easy-python-numpy-cuda-cublas/|archive-date=2009-04-20|url-status=dead|access-date=2017-08-08}}</ref>

<syntaxhighlight lang="numpy">
import numpy
from pycublas import CUBLASMatrix

A = CUBLASMatrix(numpy.mat(, ], numpy.float32))
B = CUBLASMatrix(numpy.mat(, , ], numpy.float32))
C = A * B
print(C.np_mat())
</syntaxhighlight>

while ] directly replaces NumPy:<ref>{{Cite web|url=https://cupy.dev/|title=CuPy|language=en|access-date=2020-01-08}}</ref>

<syntaxhighlight lang="numpy">
import cupy

a = cupy.random.randn(400)
b = cupy.random.randn(400)

dest = cupy.zeros_like(a)

print(dest - a * b)
</syntaxhighlight>

==GPUs supported==
Supported CUDA compute capability versions for CUDA SDK version and microarchitecture (by code name):
{| class="wikitable"
|+ Compute capability (CUDA SDK support vs. microarchitecture)
|-
! CUDA SDK<br>version(s) !! ] !! ] !! ]<br>(early) !! ]<br>(late) !! ] !! ] !! ] !! ] !! ] !! ] !! ] !! ]
|-
| 1.0<ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/1.0/NVIDIA_CUDA_Programming_Guide_1.0.pdf|title=NVIDIA CUDA Programming Guide. Version 1.0|date=June 23, 2007}}</ref> || {{yes|1.0 – 1.1}} || || || || || || || || || || ||
|-
| 1.1 || {{yes|1.0 – 1.1+x}} || || || || || || || || || || ||
|-
| 2.0 || {{yes|1.0 – 1.1+x}} || || || || || || || || || || ||
|-
| 2.1 – 2.3.1<ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/2_1/toolkit/docs/NVIDIA_CUDA_Programming_Guide_2.1.pdf|title=NVIDIA CUDA Programming Guide. Version 2.1|date=December 8, 2008}}</ref><ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/2_2/toolkit/docs/NVIDIA_CUDA_Programming_Guide_2.2.pdf|title=NVIDIA CUDA Programming Guide. Version 2.2|date=April 2, 2009}}</ref><ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/2_21/toolkit/docs/NVIDIA_CUDA_Programming_Guide_2.2.1.pdf|title=NVIDIA CUDA Programming Guide. Version 2.2.1|date=May 26, 2009}}</ref><ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/2_3/toolkit/docs/NVIDIA_CUDA_Programming_Guide_2.3.pdf|title=NVIDIA CUDA Programming Guide. Version 2.3.1|date=August 26, 2009}}</ref> || {{yes|1.0 – 1.3}} || || || || || || || || || || ||
|-
| 3.0 – 3.1<ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/3_0/toolkit/docs/NVIDIA_CUDA_ProgrammingGuide.pdf|title=NVIDIA CUDA Programming Guide. Version 3.0|date=February 20, 2010}}</ref><ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/3_1/toolkit/docs/NVIDIA_CUDA_C_ProgrammingGuide_3.1.pdf|title=NVIDIA CUDA C Programming Guide. Version 3.1.1|date=July 21, 2010}}</ref> || {{yes|1.0}} || {{yes|2.0}} || || || || || || || || || ||
|-
| 3.2<ref>{{cite web|url=http://developer.download.nvidia.com/compute/cuda/3_2_prod/toolkit/docs/CUDA_C_Programming_Guide.pdf|title=NVIDIA CUDA C Programming Guide. Version 3.2|date=November 9, 2010}}</ref> || {{yes|1.0}} || {{yes|2.1}} || || || || || || || || || ||
|-
| 4.0 – 4.2 || {{yes|1.0}} || {{yes|2.1}} || || || || || || || || || ||
|-
| 5.0 – 5.5 || {{yes|1.0}} || {{yes|}} || {{yes|}} || {{yes|3.5}} || || || || || || || ||
|-
| 6.0 || {{yes|1.0}} || {{yes|}} || {{yes|3.2}} || {{yes|3.5}} || || || || || || || ||
|-
| 6.5 || {{yes|1.1}} || {{yes|}} || {{yes|}} || {{yes|3.7}} || {{yes|5.x}} || || || || || || ||
|-
| 7.0 – 7.5 || || {{yes|2.0}} || {{yes|}} || {{yes|}} || {{yes|5.x}} || || || || || || ||
|-
| 8.0 || || {{yes|2.0}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|6.x}} || || || || || ||
|-
| 9.0 – 9.2 || || || {{yes|3.0}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|7.0 – 7.2}} || || || || ||
|-
| 10.0 – 10.2 || || || {{yes|3.0}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|7.5}} || || || ||
|-
| 11.0<ref>{{cite web|url=https://docs.nvidia.com/cuda/archive/11.0/cuda-toolkit-release-notes/index.html|title=CUDA 11.0 Release Notes|website=NVIDIA Developer}}</ref> || || || || {{yes|3.5}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|8.0}} || || ||
|-
| 11.1 – 11.4<ref>{{cite web|url=https://docs.nvidia.com/cuda/archive/11.1.0/cuda-toolkit-release-notes/index.html|title=CUDA 11.1 Release Notes|website=NVIDIA Developer}}</ref> || || || || {{yes|3.5}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|8.6}} || || ||
|-
| 11.5 – 11.7.1<ref>{{cite web|url=https://docs.nvidia.com/cuda/archive/11.5.0/cuda-toolkit-release-notes/index.html|title=CUDA 11.5 Release Notes|website=NVIDIA Developer}}</ref> || || || || {{yes|3.5}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|8.7}} || || ||
|-
| 11.8<ref>{{cite web|url=https://docs.nvidia.com/cuda/archive/11.8.0/cuda-toolkit-release-notes/index.html|title=CUDA 11.8 Release Notes|website=NVIDIA Developer}}</ref> || || || || {{yes|3.5}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|8.9}} || {{yes|9.0}} ||
|-
| 12.0 – 12.5 || || || || || {{yes|5.0}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|}} || {{yes|9.0}} ||
|}

Note: CUDA SDK 10.2 is the last official release for macOS, as support will not be available for macOS in newer releases.


CUDA compute capability by version with associated GPU semiconductors and GPU card models (separated by their various application areas):
{| class="wikitable" style="font-size: 85%; text-align: center; width: auto;" {| class="wikitable" style="font-size: 85%; text-align: center; width: auto;"
|+ Compute capability, GPU semiconductors and Nvidia GPU board products
|- |-
! Compute<br />capability<br />(version) ! Compute<br />capability<br />(version)
! ] ! ]
! GPUs ! GPUs
! ]
! Cards
! ], ]
! ]
! ],<br />],<br />]
|- |-
| 1.0 | 1.0
| rowspan="4" | ] | rowspan="4" | ]
| G80
| G80, G92, G92b, G94, G94b
|GeForce 8800 Ultra, GeForce 8800 GTX, GeForce 8800 GTS(G80)
| GeForce GT 420*, GeForce 8800 Ultra, GeForce 8800 GTX, GeForce GT 340*, GeForce GT 330*, GeForce GT 320*, GeForce 315*, GeForce 310*, GeForce 9800 GT, GeForce 9600 GT, GeForce 9400GT, Quadro FX 5600, Quadro FX 4600, Quadro Plex 2100 S4, Tesla C870, Tesla D870, Tesla S870
|Quadro FX 5600, Quadro FX 4600, Quadro Plex 2100 S4
|Tesla C870, Tesla D870, Tesla S870
|
|- |-
| 1.1 | 1.1
|G86, G84, G98, G96, G96b, G94, G94b, G92, G92b |G92, G94, G96, G98, G84, G86
|GeForce G110M, GeForce 9300M GS, GeForce 9200M GS, GeForce 9100M G, GeForce 8400M GT, GeForce 8600 GT, GeForce 8600 GTS, GeForce G105M, Quadro FX 4700 X2, Quadro FX 3700, Quadro FX 1800, Quadro FX 1700, Quadro FX 580, Quadro FX 570, Quadro FX 470, Quadro FX 380, Quadro FX 370, Quadro FX 370 Low Profile, Quadro NVS 450, Quadro NVS 420, Quadro NVS 290, Quadro NVS 295, Quadro Plex 2100 D4, Quadro FX 3800M, Quadro FX 3700M, Quadro FX 3600M, Quadro FX 2800M, Quadro FX 2700M, Quadro FX 1700M, Quadro FX 1600M, Quadro FX 770M, Quadro FX 570M, Quadro FX 370M, Quadro FX 360M, Quadro NVS 320M, Quadro NVS 160M, Quadro NVS 150M, Quadro NVS 140M, Quadro NVS 135M, Quadro NVS 130M, Quadro NVS 450, Quadro NVS 420, Quadro NVS 295 |GeForce GTS 250, GeForce 9800 GX2, GeForce 9800 GTX, GeForce 9800 GT, GeForce 8800 GTS(G92), GeForce 8800 GT, GeForce 9600 GT, GeForce 9500 GT, GeForce 9400 GT, GeForce 8600 GTS, GeForce 8600 GT, GeForce 8500 GT,<br />GeForce G110M, GeForce 9300M GS, GeForce 9200M GS, GeForce 9100M G, GeForce 8400M GT, GeForce G105M
|Quadro FX 4700 X2, Quadro FX 3700, Quadro FX 1800, Quadro FX 1700, Quadro FX 580, Quadro FX 570, Quadro FX 470, Quadro FX 380, Quadro FX 370, Quadro FX 370 Low Profile, Quadro NVS 450, Quadro NVS 420, Quadro NVS 290, Quadro NVS 295, Quadro Plex 2100 D4,<br />Quadro FX 3800M, Quadro FX 3700M, Quadro FX 3600M, Quadro FX 2800M, Quadro FX 2700M, Quadro FX 1700M, Quadro FX 1600M, Quadro FX 770M, Quadro FX 570M, Quadro FX 370M, Quadro FX 360M, Quadro NVS 320M, Quadro NVS 160M, Quadro NVS 150M, Quadro NVS 140M, Quadro NVS 135M, Quadro NVS 130M, Quadro NVS 450, Quadro NVS 420,<ref>{{cite web|url=https://www.techpowerup.com/gpu-specs/quadro-nvs-420.c1448|title=NVIDIA Quadro NVS 420 Specs|website=TechPowerUp GPU Database|date=25 August 2023 }}</ref> Quadro NVS 295
|
|
|- |-
|1.2 |1.2
|GT218, GT216, GT215 |GT218, GT216, GT215
|GeForce GT 240, GeForce GT 220*, GeForce 210*, GeForce GTS 360M, GeForce GTS 350M, GeForce GT 335M, GeForce GT 330M, GeForce GT 325M, GeForce GT 240M, GeForce G210M, GeForce 310M, GeForce 305M, Quadro FX 380 Low Profile, NVIDIA NVS 300, Quadro FX 1800M, Quadro FX 880M, Quadro FX 380M, NVIDIA NVS 300, NVS 5100M, NVS 3100M, NVS 2100M, ION |GeForce GT 340*, GeForce GT 330*, GeForce GT 320*, GeForce 315*, GeForce 310*, GeForce GT 240, GeForce GT 220, GeForce 210,<br />GeForce GTS 360M, GeForce GTS 350M, GeForce GT 335M, GeForce GT 330M, GeForce GT 325M, GeForce GT 240M, GeForce G210M, GeForce 310M, GeForce 305M
|Quadro FX 380 Low Profile, Quadro FX 1800M, Quadro FX 880M, Quadro FX 380M,<br />Nvidia NVS 300, NVS 5100M, NVS 3100M, NVS 2100M, ION
|
|
|- |-
|1.3 |1.3
|GT200, GT200b |GT200, GT200b
|GeForce GTX 280, GeForce GTX 275, GeForce GTX 260, Quadro FX 5800, Quadro FX 4800, Quadro FX 4800 for Mac, Quadro FX 3800, Quadro CX, Quadro Plex 2200 D2, Tesla C1060, Tesla S1070, Tesla M1060 |GeForce GTX 295, GTX 285, GTX 280, GeForce GTX 275, GeForce GTX 260
|Quadro FX 5800, Quadro FX 4800, Quadro FX 4800 for Mac, Quadro FX 3800, Quadro CX, Quadro Plex 2200 D2
|Tesla C1060, Tesla S1070, Tesla M1060
|
|- |-
|2.0 |2.0
| rowspan="2" | ] | rowspan="2" | ]
|GF100, GF110 |GF100, GF110
|GeForce GTX 590, GeForce GTX 580, GeForce GTX 570, GeForce GTX 480, GeForce GTX 470, GeForce GTX 465, GeForce GTX 480M, Quadro 6000, Quadro 5000, Quadro 4000, Quadro 4000 for Mac, Quadro Plex 7000, Quadro 5010M, Quadro 5000M, Tesla C2075, Tesla C2050/C2070, Tesla M2050/M2070/M2075/M2090 |GeForce GTX 590, GeForce GTX 580, GeForce GTX 570, GeForce GTX 480, GeForce GTX 470, GeForce GTX 465,<br />GeForce GTX 480M
|Quadro 6000, Quadro 5000, Quadro 4000, Quadro 4000 for Mac, Quadro Plex 7000,<br />Quadro 5010M, Quadro 5000M
|Tesla C2075, Tesla C2050/C2070, Tesla M2050/M2070/M2075/M2090
|
|- |-
|2.1 |2.1
|GF104, GF106 GF108，GF114, GF116, GF119 |GF104, GF106 GF108, GF114, GF116, GF117, GF119
|GeForce GTX 560 Ti, GeForce GTX 550 Ti, GeForce GTX 460, GeForce GTS 450, GeForce GTS 450*, GeForce GT 640 (GDDR3), GeForce GT 630, GeForce GT 620, GeForce GT 610, GeForce GT 520, GeForce GT 440, GeForce GT 440*, GeForce GT 430, GeForce GT 430*, GeForce GTX 675M, GeForce GTX 670M, GeForce GT 635M, GeForce GT 630M, GeForce GT 625M, GeForce GT 720M, GeForce GT 620M, GeForce 710M, GeForce 610M, GeForce GTX 580M, GeForce GTX 570M, GeForce GTX 560M, GeForce GT 555M, GeForce GT 550M, GeForce GT 540M, GeForce GT 525M, GeForce GT 520MX, GeForce GT 520M, GeForce GTX 485M, GeForce GTX 470M, GeForce GTX 460M, GeForce GT 445M, GeForce GT 435M, GeForce GT 420M, GeForce GT 415M, GeForce 710M, GeForce 410M, Quadro 2000, Quadro 2000D, Quadro 600, Quadro 410, Quadro 4000M, Quadro 3000M, Quadro 2000M, Quadro 1000M, NVS 5400M, NVS 5200M, NVS 4200M |GeForce GTX 560 Ti, GeForce GTX 550 Ti, GeForce GTX 460, GeForce GTS 450, GeForce GTS 450*, GeForce GT 640 (GDDR3), GeForce GT 630, GeForce GT 620, GeForce GT 610, GeForce GT 520, GeForce GT 440, GeForce GT 440*, GeForce GT 430, GeForce GT 430*, GeForce GT 420*,<br />GeForce GTX 675M, GeForce GTX 670M, GeForce GT 635M, GeForce GT 630M, GeForce GT 625M, GeForce GT 720M, GeForce GT 620M, GeForce 710M, GeForce 610M, GeForce 820M, GeForce GTX 580M, GeForce GTX 570M, GeForce GTX 560M, GeForce GT 555M, GeForce GT 550M, GeForce GT 540M, GeForce GT 525M, GeForce GT 520MX, GeForce GT 520M, GeForce GTX 485M, GeForce GTX 470M, GeForce GTX 460M, GeForce GT 445M, GeForce GT 435M, GeForce GT 420M, GeForce GT 415M, GeForce 710M, GeForce 410M
|Quadro 2000, Quadro 2000D, Quadro 600,<br />Quadro 4000M, Quadro 3000M, Quadro 2000M, Quadro 1000M,<br />NVS 310, NVS 315, NVS 5400M, NVS 5200M, NVS 4200M
|
|
|- |-
|3.0 |3.0
| rowspan="4" | ] | rowspan="4" | ]
|GK104, GK106, GK107 |GK104, GK106, GK107
|GeForce GTX 770, GeForce GTX 760, GeForce GT 740, GeForce GTX 690, GeForce GTX 680, GeForce GTX 670, GeForce GTX 660 Ti, GeForce GTX 660, GeForce GTX 650 Ti BOOST, GeForce GTX 650 Ti, GeForce GTX 650, GeForce GTX 880M, GeForce GTX 780M, GeForce GTX 770M, GeForce GTX 765M, GeForce GTX 760M, GeForce GTX 680MX, GeForce GTX 680M, GeForce GTX 675MX, GeForce GTX 670MX, GeForce GTX 660M, GeForce GT 750M, GeForce GT 650M, GeForce GT 745M, GeForce GT 645M, GeForce GT 740M, GeForce GT 730M, GeForce GT 640M, GeForce GT 640M LE, GeForce GT 735M, GeForce GT 730M, Quadro K5000, Quadro K4200, Quadro K4000, Quadro K2000, Quadro K2000D, Quadro K600, Quadro K420, Quadro K500M, Quadro K510M, Quadro K610M, Quadro K1000M, Quadro K2000M, Quadro K1100M, Quadro K2100M, Quadro K3000M, Quadro K3100M, Quadro K4000M, Quadro K5000M, Quadro K4100M, Quadro K5100M, Tesla K10 |GeForce GTX 770, GeForce GTX 760, GeForce GT 740, GeForce GTX 690, GeForce GTX 680, GeForce GTX 670, GeForce GTX 660 Ti, GeForce GTX 660, GeForce GTX 650 Ti BOOST, GeForce GTX 650 Ti, GeForce GTX 650,<br />GeForce GTX 880M, GeForce GTX 870M, GeForce GTX 780M, GeForce GTX 770M, GeForce GTX 765M, GeForce GTX 760M, GeForce GTX 680MX, GeForce GTX 680M, GeForce GTX 675MX, GeForce GTX 670MX, GeForce GTX 660M, GeForce GT 750M, GeForce GT 650M, GeForce GT 745M, GeForce GT 645M, GeForce GT 740M, GeForce GT 730M, GeForce GT 640M, GeForce GT 640M LE, GeForce GT 735M, GeForce GT 730M
|Quadro K5000, Quadro K4200, Quadro K4000, Quadro K2000, Quadro K2000D, Quadro K600, Quadro K420,<br />Quadro K500M, Quadro K510M, Quadro K610M, Quadro K1000M, Quadro K2000M, Quadro K1100M, Quadro K2100M, Quadro K3000M, Quadro K3100M, Quadro K4000M, Quadro K5000M, Quadro K4100M, Quadro K5100M,<br />NVS 510, Quadro 410
|Tesla K10, GRID K340, GRID K520, GRID K2
|
|- |-
|3.2 |3.2
|GK20A
|Tegra K1
|
|Jetson TK1 (SoC)
|
|
|Tegra&nbsp;K1,<br />Jetson&nbsp;TK1
|- |-
|3.5 |3.5
|GK110, GK208 |GK110, GK208
|GeForce GTX TITAN Z, GeForce GTX TITAN Black, GeForce GTX TITAN, GeForce GTX 780 Ti, GeForce GTX 780, GeForce GT 640 (GDDR5), GeForce GT 630 v2, GeForce GT 730, GeForce GT 720, Quadro K6000, Quadro K5200, Tesla K40, Tesla K20x, Tesla K20 |GeForce GTX Titan Z, GeForce GTX Titan Black, GeForce GTX Titan, GeForce GTX 780 Ti, GeForce GTX 780, GeForce GT 640 (GDDR5), GeForce GT 630 v2, GeForce GT 730, GeForce GT 720, GeForce GT 710, GeForce GT 740M (64-bit, DDR3), GeForce GT 920M
|Quadro K6000, Quadro K5200
|Tesla K40, Tesla K20x, Tesla K20
|
|- |-
|3.7 |3.7
|GK210 |GK210
|
|
| Tesla K80 | Tesla K80
|
|- |-
|5.0 |5.0
| rowspan="2" | ] | rowspan="3" | ]
|GM107, GM108 |GM107, GM108
|GeForce GTX 750 Ti, GeForce GTX 750, GeForce GTX 960M, GeForce GTX 950M, GeForce 940M, GeForce 930M, GeForce GTX 860M, GeForce GTX 850M, GeForce 845M, GeForce 840M, GeForce 830M, Quadro K2200, Quadro K1200, Quadro K620, Quadro K620M |GeForce GTX 750 Ti, GeForce GTX 750, GeForce GTX 960M, GeForce GTX 950M, GeForce 940M, GeForce 930M, GeForce GTX 860M, GeForce GTX 850M, GeForce 845M, GeForce 840M, GeForce 830M
|Quadro K1200, Quadro K2200, Quadro K620, Quadro M2000M, Quadro M1000M, Quadro M600M, Quadro K620M, NVS 810
|Tesla M10
|
|- |-
|5.2 |5.2
|GM200, GM204, GM206 |GM200, GM204, GM206
| GeForce GTX TITAN X, GeForce GTX 980, GeForce GTX 970, GeForce GTX 960, GeForce GTX 980M, GeForce GTX 970M, GeForce GTX 965M, Quadro M6000 |GeForce GTX Titan X, GeForce GTX 980 Ti, GeForce GTX 980, GeForce GTX 970, GeForce GTX 960, GeForce GTX 950, GeForce GTX 750 SE,<br />GeForce GTX 980M, GeForce GTX 970M, GeForce GTX 965M
|Quadro M6000 24GB, Quadro M6000, Quadro M5000, Quadro M4000, Quadro M2000, Quadro M5500,<br />Quadro M5000M, Quadro M4000M, Quadro M3000M
|Tesla M4, Tesla M40, Tesla M6, Tesla M60
|
|-
|5.3
|GM20B
|
|
|
|Tegra&nbsp;X1,<br />Jetson&nbsp;TX1,<br />Jetson&nbsp;Nano,<br />DRIVE&nbsp;CX,<br />DRIVE&nbsp;PX
|-
|6.0
| rowspan="3" |]
|GP100
|
| Quadro GP100
| Tesla P100
|
|-
|6.1
|GP102, GP104, GP106, GP107, GP108
|Nvidia TITAN Xp, Titan X,<br />GeForce GTX 1080 Ti, GTX 1080, GTX 1070 Ti, GTX 1070, GTX 1060,<br /> GTX 1050 Ti, GTX 1050, GT 1030, GT 1010,<br /> MX350, MX330, MX250, MX230, MX150, MX130, MX110
|Quadro P6000, Quadro P5000, Quadro P4000, Quadro P2200, Quadro P2000, Quadro P1000, Quadro P400, Quadro P500, Quadro P520, Quadro P600,<br />Quadro P5000 (mobile), Quadro P4000 (mobile), Quadro P3000 (mobile)
|Tesla P40, Tesla P6, Tesla P4
|
|-
|6.2
|GP10B<ref>{{cite web|url=http://www.phoronix.com/scan.php?page=news_item&px=Tegra-X2-Nouveau-Support|title=NVIDIA Rolls Out Tegra X2 GPU Support In Nouveau|last=Larabel|first=Michael|author-link=Michael Larabel|publisher=]|date=March 29, 2017|access-date=August 8, 2017}}</ref>
|
|
|
|Tegra&nbsp;X2, Jetson&nbsp;TX2, DRIVE&nbsp;PX&nbsp;2
|-
|7.0
| rowspan="2" |]
|GV100
|NVIDIA TITAN V
|Quadro GV100
|Tesla V100, Tesla V100S
|
|-
|7.2
|GV10B<ref> on TechPowerUp (preliminary)</ref><br />
GV11B<ref>{{Cite web | url=https://docs.nvidia.com/jetson/l4t/index.html#page/Tegra%20Linux%20Driver%20Package%20Development%20Guide/power_management_jetson_xavier.html | title=Welcome — Jetson LinuxDeveloper Guide 34.1 documentation }}</ref><ref>{{Cite web | url=https://www.phoronix.com/scan.php?page=news_item&px=NVIDIA-Nouveau-GV11B-Volta-Xav | title=NVIDIA Bringing up Open-Source Volta GPU Support for Their Xavier SoC }}</ref>
|
|
|
|Tegra Xavier,<br />Jetson Xavier NX,<br />Jetson AGX Xavier,<br />DRIVE AGX Xavier,<br />DRIVE AGX Pegasus,<br />Clara AGX
|-
|7.5
||]
|TU102, TU104, TU106, TU116, TU117
|NVIDIA TITAN RTX,<br />GeForce RTX 2080 Ti, RTX 2080 Super, RTX 2080, RTX 2070 Super, RTX 2070, RTX 2060 Super, RTX 2060 12GB, RTX 2060,<br />GeForce GTX 1660 Ti, GTX 1660 Super, GTX 1660, GTX 1650 Super, GTX 1650, MX550, MX450
|Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 5000, Quadro RTX 4000, T1000, T600, T400<br />T1200 (mobile), T600 (mobile), T500 (mobile), Quadro T2000 (mobile), Quadro T1000 (mobile)
|Tesla T4
|
|-
|8.0
| rowspan="3" |]
|GA100
|
|
|A100 80GB, A100 40GB, A30
|
|-
|8.6
|GA102, GA103, GA104, GA106, GA107
|GeForce RTX 3090 Ti, RTX 3090, RTX 3080 Ti, RTX 3080 12GB, RTX 3080, RTX 3070 Ti, RTX 3070, RTX 3060 Ti, RTX 3060, RTX 3050, RTX 3050 Ti (mobile), RTX 3050 (mobile), RTX 2050 (mobile), MX570
|RTX A6000, RTX A5500, RTX A5000, RTX A4500, RTX A4000, RTX A2000<br /> RTX A5000 (mobile), RTX A4000 (mobile), RTX A3000 (mobile), RTX A2000 (mobile)
|A40, A16, A10, A2
|
|-
|8.7
|GA10B
|
|
|
|Jetson Orin Nano,<br />Jetson Orin NX,<br />Jetson AGX Orin,<br />DRIVE AGX Orin,<br />DRIVE&nbsp;AGX&nbsp;Pegasus&nbsp;OA,<br />Clara Holoscan
|-
|8.9
|]<ref>{{cite web | url=https://www.nvidia.com/en-us/geforce/ada-lovelace-architecture/ | title=NVIDIA Ada Lovelace Architecture }}</ref>
|AD102, AD103, AD104, AD106, AD107
|GeForce RTX 4090, RTX 4080 Super, RTX 4080, RTX 4070 Ti Super, RTX 4070 Ti, RTX 4070 Super, RTX 4070, RTX 4060 Ti, RTX 4060, RTX 4050 (mobile)
|RTX 6000 Ada, RTX 5880 Ada, RTX 5000 Ada, RTX 4500 Ada, RTX 4000 Ada, RTX 4000 SFF
|L40S, L40, L20, L4, L2
|
|-
|9.0
|]
|GH100
|
|
|H200, H100
|
|-
|10.0
|rowspan="2" |]
|GB100
|
|
|B200, B100
|
|-
|10.x
|GB202, GB203, GB205, GB206, GB207
|GeForce RTX 5090, RTX 5080, RTX 5070 Ti, RTX 5070
|
|B40
|
|-
!Compute<br />capability<br />(version)
!]
!GPUs
!]
!], ]
!]
!],<br />],<br />]
|} |}
'*' - OEM-only products '*' – ]-only products


==Version features and specifications==
A table of devices officially supporting CUDA:<ref name="CUDA_products" />
{{Update|section|reason=Missing CUDA compute capability 10.x (Blackwell)|date=March 2024}}
{| class="wikitable" style="font-size:85%;"
|-
! rowspan=2 | Feature support (unlisted features are supported for all compute capabilities)
! colspan="14" | Compute capability (version)
|-
! 1.0, 1.1 !! 1.2, 1.3 !! 2.x !! 3.0 !! 3.2 !! 3.5, 3.7, 5.x, 6.x, 7.0, 7.2 !! 7.5 !! 8.x !! 9.0
|-
| Warp vote functions (__all(), __any())
| colspan="1" {{no}}
| colspan="8" {{yes}}
|-
| Warp vote functions (__ballot())
| colspan="2" rowspan="5" {{no}}
| colspan="7" rowspan="5" {{yes}}
|-
| Memory fence functions (__threadfence_system())
|-
| Synchronization functions (__syncthreads_count(), __syncthreads_and(), __syncthreads_or())
|-
| Surface functions
|-
| 3D grid of thread blocks
|-
| Warp shuffle functions
| colspan="3" rowspan="2" {{no}}
| colspan="6" rowspan="2" {{yes}}
|-
| Unified memory programming
|-
| Funnel shift
| colspan="4" rowspan="1" {{no}}
| colspan="5" rowspan="1" {{yes}}
|-
| Dynamic parallelism
| colspan="5" rowspan="1" {{no}}
| colspan="4" rowspan="1" {{yes}}
|-
| Uniform Datapath<ref></ref>
| colspan="6" rowspan="1" {{no}}
| colspan="3" rowspan="1" {{yes}}
|-
| Hardware-accelerated async-copy
| colspan="7" rowspan="4" {{no}}
| colspan="2" rowspan="4" {{yes}}
|-
| Hardware-accelerated ''split arrive/wait barrier''
|-
| Warp-level support for reduction ops
|-
| L2 cache residency management
|-
| DPX instructions for accelerated dynamic programming
| colspan="8" rowspan="4" {{no}}
| colspan="1" rowspan="4" {{yes}}
|-
| Distributed shared memory
|-
| Thread block cluster
|-
| Tensor memory accelerator (TMA) unit
|-
! rowspan="2" |Feature support (unlisted features are supported for all compute capabilities)
!1.0,1.1
!1.2,1.3
!2.x
!3.0
!3.2
!3.5, 3.7, 5.x, 6.x, 7.0, 7.2
!7.5
!8.x
!9.0
|-
! colspan="14" |Compute capability (version)
|}<ref>{{Cite web|url=https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications|title=<!--CUDA C Programming Guide-->H.1. Features and Technical Specifications{{snd}} Table 13. Feature Support per Compute Capability|website=docs.nvidia.com|language=en-us|access-date=2020-09-23}}</ref>


===Data types===
{| cellpadding="12" style="font-size: 85%;"
{| class="wikitable" style="font-size:85%;"
|- valign="top"
| |-
! Data type
{| class="standard"
! Operation
!]
! Supported since<br />
|-
! Atomic Operation
|GeForce GTX TITAN X
! Supported since<br />for global memory
|-
! Supported since<br />for shared memory
|GeForce GTX 980
|- |-
| 8-bit integer<br />signed/unsigned
|GeForce GTX 970
| loading, storing, conversion
|-
| {{yes|1.0}}
|GeForce GTX 960
| {{n/a}}
|-
| colspan="2" {{n/a}}
|GeForce GTX Titan Z
|- |-
| 16-bit integer<br />signed/unsigned
|GeForce GTX TITAN Black
| general operations
|-
| {{yes|1.0}}
|GeForce GTX TITAN
| atomicCAS()
|-
| colspan="2" {{yes|3.5}}
|GeForce GTX 780 Ti
|- |-
| 32-bit integer<br />signed/unsigned
|GeForce GTX 780
| general operations
|-
| {{yes|1.0}}
|GeForce GTX 770
| atomic functions
|-
| {{yes|1.1}}
|GeForce GTX 760
| {{yes|1.2}}
|-
|-
|GeForce GTX 750 Ti
| 64-bit integer<br />signed/unsigned
|-
| general operations
|GeForce GTX 750
| {{yes|1.0}}
|-
| atomic functions
|GeForce GT 740
| {{yes|1.2}}
|-
| {{yes|2.0}}
|GeForce GT 730
|- |-
| any 128-bit trivially copyable type
|GeForce GTX 690
| general operations
|-
| {{no}}
|GeForce GTX 680
| atomicExch, atomicCAS
|-
| colspan="2" {{yes|9.0}}
|GeForce GTX 670
|- |-
| rowspan="2" | 16-bit floating point<br />FP16
|GeForce GTX 660 Ti
| rowspan="2" | addition, subtraction,<br />multiplication, comparison,<br />warp shuffle functions, conversion
|-
| rowspan="2" {{yes|5.3}}
|GeForce GTX 660
| half2 atomic addition
|-
| colspan="2" {{yes|6.0}}
|GeForce GTX 650 Ti BOOST
|- |-
| atomic addition
|GeForce GTX 650 Ti
| colspan="2" {{yes|7.0}}
|-
|-
|GeForce GTX 650
| 16-bit floating point<br />BF16
|-
| addition, subtraction,<br />multiplication, comparison,<br />warp shuffle functions, conversion
|GeForce GT 640
| {{yes|8.0}}
|-
| atomic addition
|GeForce GT 630
| colspan="2" {{yes|8.0}}
|-
|-
|GeForce GT 620
| rowspan="2" | 32-bit floating point
|-
| rowspan="2" | general operations
|GeForce GT 610
| rowspan="2" {{yes|1.0}}
|-
| atomicExch()
|GeForce GTX 590
| {{yes|1.1}}
|-
| {{yes|1.2}}
|GeForce GTX 580
|- |-
| atomic addition
|GeForce GTX 570
| colspan="2" {{yes|2.0}}
|-
|-
|GeForce GTX 560 Ti
| rowspan="1" | 32-bit floating point float2 and float4
|-
| general operations
|GeForce GTX 560
| {{no}}
|-
| atomic addition
|GeForce GTX 550 Ti
| colspan="2" {{yes|9.0}}
|-
|-
|GeForce GT 520
| rowspan="1" | 64-bit floating point
|-
| general operations
|GeForce GTX 480
| {{yes|1.3}}
|-
| atomic addition
|GeForce GTX 470
| colspan="2" {{yes|6.0}}
|-
|GeForce GTX 465
|-
|GeForce GTX 460
|-
|GeForce GTX 460 SE
|-
|GeForce GTS 450
|-
|GeForce GT 440
|-
|GeForce GT 430
|-
|GeForce GT 420
|-
|GeForce GTX 295
|-
|GeForce GTX 285
|-
|GeForce GTX 280
|-
|GeForce GTX 275
|-
|GeForce GTX 260
|-
|GeForce GTS 250
|-
|GeForce GTS 240
|-
|GeForce GT 240
|-
|GeForce GT 220
|-
|GeForce 210/G210
|-
|GeForce GT 140
|-
|GeForce 9800 GX2
|-
|GeForce 9800 GTX+
|-
|GeForce 9800 GTX
|-
|GeForce 9800 GT
|-
|GeForce 9600 GSO
|-
|GeForce 9600 GT
|-
|GeForce 9500 GT
|-
|GeForce 9400 GT
|-
|GeForce 9400 mGPU
|-
|GeForce 9300 mGPU
|-
|GeForce 9100 mGPU
|-
|GeForce 8800 Ultra
|-
|GeForce 8800 GTX
|-
|GeForce 8800 GTS
|-
|GeForce 8800 GT
|-
|GeForce 8800 GS
|-
|GeForce 8600 GTS
|-
|GeForce 8600 GT
|-
|GeForce 8600 mGT
|-
|GeForce 8500 GT
|-
|GeForce 8400 GS
|-
|GeForce 8300 mGPU
|-
|GeForce 8200 mGPU
|-
|GeForce 8100 mGPU
|}

|
{| class="standard"
!]
|-
|GeForce GTX 980M
|-
|GeForce GTX 970M
|-
|GeForce GTX 965M
|-
|GeForce GTX 960M
|-
|GeForce GTX 950M
|-
|GeForce 940M
|-
|GeForce 930M
|-
|GeForce GTX 880M
|-
|GeForce GTX 870M
|-
|GeForce GTX 860M
|-
|GeForce GTX 850M
|-
|GeForce 845M
|-
|GeForce 840M
|-
|GeForce 830M
|-
|GeForce GTX 780M
|-
|GeForce GTX 770M
|-
|GeForce GTX 765M
|-
|GeForce GTX 760M
|-
|GeForce GT 750M
|-
|GeForce GT 745M
|-
|GeForce GT 740M
|-
|GeForce GT 735M
|-
|GeForce GT 730M
|-
|GeForce GTX 680MX
|-
|GeForce GTX 680M
|-
|GeForce GTX 675MX
|-
|GeForce GTX 675M
|-
|GeForce GTX 670MX
|-
|GeForce GTX 670M
|-
|GeForce GTX 660M
|-
|GeForce GT 650M
|-
|GeForce GT 645M
|-
|GeForce GT 640M
|-
|GeForce GTX 580M
|-
|GeForce GTX 570M
|-
|GeForce GTX 560M
|-
|GeForce GT 555M
|-
|GeForce GT 550M
|-
|GeForce GT 540M
|-
|GeForce GT 525M
|-
|GeForce GT 520M
|-
|GeForce GTX 480M
|-
|GeForce GTX 470M
|-
|GeForce GTX 460M
|-
|GeForce GT 445M
|-
|GeForce GT 435M
|-
|GeForce GT 425M
|-
|GeForce GT 420M
|-
|GeForce GT 415M
|-
|GeForce GTX 285M
|-
|GeForce GTX 280M
|-
|GeForce GTX 260M
|-
|GeForce GTS 360M
|-
|GeForce GTS 350M
|-
|GeForce GTS 260M
|-
|GeForce GTS 250M
|-
|GeForce GT 335M
|-
|GeForce GT 330M
|-
|GeForce GT 325M
|-
|GeForce GT 320M
|-
|GeForce 310M
|-
|GeForce GT 240M
|-
|GeForce GT 230M
|-
|GeForce GT 220M
|-
|GeForce G210M
|-
|GeForce GTS 160M
|-
|GeForce GTS 150M
|-
|GeForce GT 130M
|-
|GeForce GT 120M
|-
|GeForce G110M
|-
|GeForce G105M
|-
|GeForce G103M
|-
|GeForce G102M
|-
|GeForce G100
|-
|GeForce 9800M GTX
|-
|GeForce 9800M GTS
|-
|GeForce 9800M GT
|-
|GeForce 9800M GS
|-
|GeForce 9700M GTS
|-
|GeForce 9700M GT
|-
|GeForce 9650M GT
|-
|GeForce 9650M GS
|-
|GeForce 9600M GT
|-
|GeForce 9600M GS
|-
|GeForce 9500M GS
|-
|GeForce 9500M G
|-
|GeForce 9400M G
|-
|GeForce 9300M GS
|-
|GeForce 9300M G
|-
|GeForce 9200M GS
|-
|GeForce 9100M G
|-
|GeForce 8800M GTX
|-
|GeForce 8800M GTS
|-
|GeForce 8700M GT
|-
|GeForce 8600M GT
|-
|GeForce 8600M GS
|-
|GeForce 8400M GT
|-
|GeForce 8400M GS
|-
|GeForce 8400M G
|-
|GeForce 8200M G
|}
|
{| class="standard"
!]
|-
|Quadro M6000
|-
|Quadro K6000
|-
|Quadro K5200
|-
|Quadro K5000
|-
|Quadro K4200
|-
|Quadro K4000
|-
|Quadro K2200
|-
|Quadro K2000D
|-
|Quadro K2000
|-
|Quadro K1200
|-
|Quadro K620
|-
|Quadro K600
|-
|Quadro K420
|-
|Quadro 6000
|-
|Quadro 5000
|-
|Quadro 4000
|-
|Quadro 2000
|-
|Quadro 600
|-
|Quadro FX 5800
|-
|Quadro FX 5600
|-
|Quadro FX 4800
|-
|Quadro FX 4700 X2
|-
|Quadro FX 4600
|-
|Quadro FX 3800
|-
|Quadro FX 3700
|-
|Quadro FX 1800
|-
|Quadro FX 1700
|-
|Quadro FX 580
|-
|Quadro FX 570
|-
|Quadro FX 380
|-
|Quadro FX 370
|-
|Quadro NVS 510
|-
|Quadro NVS 450
|-
|Quadro NVS 420
|-
|Quadro NVS 295
|-
|Quadro Plex 1000 Model IV
|-
|Quadro Plex 1000 Model S4
|}
{| class="standard"
!]
|-
|Quadro K5100M
|-
|Quadro K5000M
|-
|Quadro K4100M
|-
|Quadro K4000M
|-
|Quadro K3100M
|-
|Quadro K3000M
|-
|Quadro K2100M
|-
|Quadro K2000M
|-
|Quadro K1100M
|-
|Quadro K1000M
|-
|Quadro K620M
|-
|Quadro K610M
|-
|Quadro K510M
|-
|Quadro K500M
|-
|Quadro 5010M
|-
|Quadro 5000M
|-
|Quadro 4000M
|-
|Quadro 3000M
|-
|Quadro 2000M
|-
|Quadro 1000M
|-
|Quadro FX 3800M
|-
|Quadro FX 3700M
|-
|Quadro FX 3600M
|-
|Quadro FX 2800M
|-
|Quadro FX 2700M
|-
|Quadro FX 1800M
|-
|Quadro FX 1700M
|-
|Quadro FX 1600M
|-
|Quadro FX 880M
|-
|Quadro FX 770M
|-
|Quadro FX 570M
|-
|Quadro FX 380M
|-
|Quadro FX 370M
|-
|Quadro FX 360M
|-
|Quadro NVS 320M
|-
|Quadro NVS 160M
|-
|Quadro NVS 150M
|-
|Quadro NVS 140M
|-
|Quadro NVS 135M
|-
|Quadro NVS 130M
|}
{| class="standard"
!]
|-
|Tesla K80
|-
|Tesla K40
|-
|Tesla K20X
|-
|Tesla K20
|-
|Tesla K10
|-
|Tesla C2050/2070
|-
|Tesla M2050/M2070
|-
|Tesla S2050
|-
|Tesla S1070
|-
|Tesla M1060
|-
|Tesla C1060
|-
|Tesla C870
|-
|Tesla D870
|-
|Tesla S870
|}
|} |}
Note: Any missing lines or empty entries do reflect some lack of information on that exact item.<ref>{{cite web | url=https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications | title=CUDA C++ Programming Guide }}</ref>


===Tensor cores===
==Version features and specifications==
{| class="wikitable" style="font-size: 85%; text-align: center; width: auto;" {| class="wikitable" style="font-size:85%;"
|- |-
! FMA per cycle per tensor core<ref>Fused-Multiply-Add, actually executed, Dense Matrix</ref>
! rowspan=2 | Feature support (unlisted features are<br />supported for all compute capabilities)
! colspan=10 | Compute capability (version) ! colspan="2" | Supported since
! 7.0
! 7.2
! 7.5 Workstation
! 7.5 Desktop
! 8.0
! 8.6 Workstation
! 8.7
! 8.6 Desktop
! 8.9 Desktop
! 8.9 Workstation
! 9.0
! 10.0
|- |-
! Data Type
! 1.0 !! 1.1 !! 1.2 !! 1.3 !! 2.x !! 3.0 !! 3.5 !! 3.7 !! 5.0 !! 5.2
! For dense matrices
! For sparse matrices
| {{n/a|1st Gen (8x/SM)}}
| {{n/a|1st Gen? (8x/SM)}}
| colspan="2" {{n/a|2nd Gen (8x/SM)}}
| colspan="4" {{n/a|3rd Gen (4x/SM)}}
| colspan="3" {{n/a|4th Gen (4x/SM)}}
| colspan="1" {{n/a|5th Gen (4x/SM)}}
|- |-
| 1-bit values (AND)
! style="text-align:left;"| Integer atomic functions operating on<br />32-bit words in global memory
| {{yes|8.0 as<br />experimental}}
| colspan="1" rowspan="2" {{no}}
| colspan="9" rowspan="2" {{yes}} | rowspan="2" {{no}}
| colspan="4" {{no}}
| colspan="3" rowspan="2" {{yes|4096}}
| colspan="3" rowspan="2" {{yes|2048}}
| colspan="2" {{yes| speed tbd}}
|- |-
| 1-bit values (XOR)
! style="text-align:left;"| atomicExch() operating on 32-bit<br />floating point values in global memory
| rowspan="2" {{maybe|7.5–8.9 as<br />experimental}}
| rowspan="2" colspan="2" {{no}}
| colspan="2" {{yes|1024}}
| rowspan="2" colspan="2" {{maybe|Deprecated or removed?}}
|- |-
| 4-bit integers
! style="text-align:left;"| Integer atomic functions operating on<br />32-bit words in shared memory
| {{maybe|8.0–8.9 as<br />experimental}}
| colspan="2" rowspan="4" {{no}}
| colspan="8" rowspan="4" {{yes}} | colspan="2" {{yes|256}}
| colspan="3" {{yes|1024}}
| colspan="3" {{yes|512}}
|- |-
! style="text-align:left;"| atomicExch() operating on 32-bit<br />floating point values in shared memory | 4-bit floating point FP4 (E2M1?)
| colspan="2" {{yes|10.0}}
| colspan="11" {{no}}
| colspan="1" {{yes|4096}}
|- |-
| 6-bit floating point FP6 (E3M2 and E2M3?)
! style="text-align:left;"| Integer atomic functions operating on<br />64-bit words in global memory
| colspan="2" {{yes|10.0}}
| colspan="11" {{no}}
| colspan="1" {{yes|2048}}
|- |-
| 8-bit integers
! style="text-align:left;"| Warp vote functions
| {{yes|7.2}}
| {{yes|8.0}}
| {{no}}
| colspan="1" {{yes|128}}
| colspan="2" {{yes|128}}
| colspan="3" {{yes|512}}
| colspan="3" {{yes|256}}
| rowspan="3" {{yes|1024}}
| rowspan="3" {{yes|2048}}
|- |-
| 8-bit floating point FP8 (E4M3 and E5M2) with FP16 accumulate
! style="text-align:left;"|Double-precision floating-point operations
| colspan="3" rowspan="1" {{no}} | rowspan="2" colspan="2" {{yes|8.9}}
| colspan="7" rowspan="1" {{yes}} | rowspan="2" colspan="8" {{no}}
| rowspan="2" colspan="2" {{yes|256}}
|- |-
| 8-bit floating point FP8 (E4M3 and E5M2) with FP32 accumulate
! style="text-align:left;"| Atomic functions operating on 64-bit<br />integer values in shared memory
| colspan="4" rowspan="7" {{no}}
| colspan="6" rowspan="7" {{yes}}
|- |-
| 16-bit floating point FP16 with FP16 accumulate
! style="text-align:left;"| Floating-point atomic addition operating on<br />32-bit words in global and shared memory
| rowspan="2" {{yes|7.0}}
| rowspan="2" {{yes|8.0}}
| rowspan="2" colspan="2" {{yes|64}}
| rowspan="2" colspan="1" {{yes|64}}
| colspan="1" {{yes|64}}
| rowspan="3" colspan="3" {{yes|256}}
| colspan="3" {{yes|128}}
| rowspan="3" {{yes|512}}
| rowspan="3" {{yes|1024}}
|- |-
| 16-bit floating point FP16 with FP32 accumulate
! style="text-align:left;"| _ballot()
| rowspan="3" {{yes|32}}
| colspan="2" rowspan="2" {{yes|64}}
| colspan="1" rowspan="2" {{yes|128}}
|- |-
| 16-bit floating point BF16 with FP32 accumulate
! style="text-align:left;"| _threadfence_system()
| rowspan="2" colspan="1" {{yes|7.5<ref>as SASS since 7.5, as PTX since 8.0</ref>}}
| rowspan="2" colspan="1" {{yes|8.0}}
| rowspan="3" colspan="2" {{no}}
| colspan="1" {{maybe|64<ref name="unofficial support in SASS">unofficial support in SASS</ref>}}
|- |-
| 32-bit (19 bits used) floating point TF32
! style="text-align:left;"| _syncthreads_count(),<br />_syncthreads_and(),<br />_syncthreads_or()
| colspan="1" {{maybe|speed tbd (32?)<ref name="unofficial support in SASS">unofficial support in SASS</ref>}}
| colspan="3" {{yes|128}}
| colspan="2" {{yes|32}}
| colspan="1" {{yes|64}}
| {{yes|256}}
| {{yes|512}}
|- |-
| 64-bit floating point
! style="text-align:left;"| Surface functions
| {{yes|8.0}}
| {{no}}
| colspan="2" {{no}}
| {{yes|16}}
| colspan="5" {{yes|speed tbd}}
| {{yes|32}}
| {{yes|16}}
|}

Note: Any missing lines or empty entries do reflect some lack of information on that exact item.<ref>{{cite web|url=https://www.nvidia.com/content/dam/en-zz/Solutions/gtcf21/jetson-orin/nvidia-jetson-agx-orin-technical-brief.pdf|title=Technical brief. NVIDIA Jetson AGX Orin Series|website=nvidia.com|access-date=5 September 2023}}</ref><ref>{{cite web|url=https://images.nvidia.com/aem-dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf|title=NVIDIA Ampere GA102 GPU Architecture|website=nvidia.com|access-date=5 September 2023}}</ref>
<ref>{{cite arXiv|title=Benchmarking and Dissecting the Nvidia Hopper GPU Architecture|eprint=2402.13499v1 |last1=Luo |first1=Weile |last2=Fan |first2=Ruibo |last3=Li |first3=Zeyu |last4=Du |first4=Dayou |last5=Wang |first5=Qiang |last6=Chu |first6=Xiaowen |date=2024 |class=cs.AR }}</ref>
<ref>{{cite web|url=https://images.nvidia.com/content/Solutions/data-center/a40/nvidia-a40-datasheet.pdf|title=Datasheet NVIDIA A40|website=nvidia.com|access-date=27 April 2024}}</ref>
<ref>{{cite web | url=https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.1.pdf | title=NVIDIA AMPERE GA102 GPU ARCHITECTURE | date=27 April 2024 }}</ref>
<ref>{{cite web | url=https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/support-guide/NVIDIA-L40-Datasheet-January-2023.pdf | title=Datasheet NVIDIA L40 | date=27 April 2024 }}</ref>

{| class="wikitable" style="font-size:85%;"
|- |-
! Tensor Core Composition
! style="text-align:left;"| 3D grid of thread block
! 7.0
! 7.2, 7.5
! 8.0, 8.6
! 8.7
! 8.9
! 9.0
|- |-
| Dot Product Unit Width in FP16 units (in bytes)<ref>In the Whitepapers the Tensor Core cube diagrams represent the Dot Product Unit Width into the height (4 FP16 for Volta and Turing, 8 FP16 for A100, 4 FP16 for GA102, 16 FP16 for GH100). The other two dimensions represent the number of Dot Product Units (4x4 = 16 for Volta and Turing, 8x4 = 32 for Ampere and Hopper). The resulting gray blocks are the FP16 FMA operations per cycle. Pascal without Tensor core is only shown for speed comparison as is Volta V100 with non-FP16 datatypes.</ref><ref>{{cite web|url=https://images.nvidia.com/aem-dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf|title=NVIDIA Turing Architecture Whitepaper
! style="text-align:left;"| Warp shuffle functions
|website=nvidia.com|access-date=5 September 2023}}</ref><ref>{{cite web|url=https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a100/pdf/nvidia-a100-datasheet-us-nvidia-1758950-r4-web.pdf|title=NVIDIA Tensor Core GPU
| colspan="5" rowspan="1" {{no}}
|website=nvidia.com|access-date=5 September 2023}}</ref><ref>{{cite web | url=https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/ | title=NVIDIA Hopper Architecture In-Depth | date=22 March 2022 }}</ref>
| colspan="5" rowspan="1" {{yes}}

| colspan="2" {{yes| 4 (8)}}
| colspan="1" {{yes| 8 (16)}}
| colspan="2" {{yes| 4 (8)}}
| {{yes| 16 (32)}}
|- |-
| Dot Product Units per Tensor Core
! style="text-align:left;"| Funnel shift
| colspan="6" rowspan="2" {{no}} | colspan="2" {{yes|16}}
| colspan="4" rowspan="2" {{yes}} | colspan="4" {{yes|32}}
|- |-
| Tensor Cores per SM partition
! style="text-align:left;"| Dynamic parallelism
| colspan="2" {{yes|2}}
| colspan="4" {{yes|1}}
|- |-
| Full throughput (Bytes/cycle)<ref name="ReferenceC">shape x converted operand size, e.g. 2 tensor cores x 4x4x4xFP16/cycle = 256 Bytes/cycle</ref> per SM partition<ref name="product first 3 table rows">= product first 3 table rows</ref>
! rowspan=2 | Feature support (unlisted features are<br />supported for all compute capabilities)
| colspan="2" {{yes|256}}
! 1.0 !! 1.1 !! 1.2 !! 1.3 !! 2.x !! 3.0 !! 3.5 !! 3.7 !! 5.0 !! 5.2
| colspan="1" {{yes|512}}
| colspan="1" {{yes|256}}
|
| colspan="1" {{yes|1024}}
|- |-
| FP Tensor Cores: Minimum cycles for warp-wide matrix calculation
! colspan=10 | Compute capability (version)
| colspan="2" {{yes|8}}
| colspan="1" {{yes|4}}
| colspan="1" {{yes|8}}
|
|
|-
| FP Tensor Cores: Minimum Matrix Shape for full throughput (Bytes)<ref name="ReferenceD">= product of previous 2 table rows; shape: e.g. 8x8x4xFP16 = 512 Bytes</ref>
| colspan="4" {{yes|2048}}
|
|
|-
| INT Tensor Cores: Minimum cycles for warp-wide matrix calculation
| {{no}}
| colspan="3" {{yes|4}}
|
|
|-
| INT Tensor Cores: Minimum Matrix Shape for full throughput (Bytes)
| {{no}}
| colspan="1" {{yes|1024}}
| colspan="1" {{yes|2048}}
| colspan="1" {{yes|1024}}
|
|
|} |}
<ref>{{cite journal | last1=Sun | first1=Wei | last2=Li | first2=Ang | last3=Geng | first3=Tong | last4=Stuijk | first4=Sander | last5=Corporaal | first5=Henk | title=Dissecting Tensor Cores via Microbenchmarks: Latency, Throughput and Numeric Behaviors | journal=IEEE Transactions on Parallel and Distributed Systems| volume=34 | issue=1 |year=2023| doi=10.1109/tpds.2022.3217824 | pages=246–261| arxiv=2206.02874 | s2cid=249431357 }}</ref><ref>{{cite web | url=https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#warp-level-matrix-instructions-mma | title=Parallel Thread Execution ISA Version 7.7 }}</ref><ref>{{cite arXiv | last1=Raihan | first1=Md Aamir | last2=Goli | first2=Negar | last3=Aamodt | first3=Tor | title=Modeling Deep Learning Accelerator Enabled GPUs | date=2018 | class=cs.MS | eprint=1811.08309 }}</ref><ref>{{cite web | url=https://www.nvidia.com/en-gb/geforce/ada-lovelace-architecture | title=NVIDIA Ada Lovelace Architecture }}</ref>


{| class="wikitable" style="font-size: 85%; text-align: center; width: auto;" {| class="wikitable" style="font-size:85%;"
|-
! FP64 Tensor Core Composition
! 8.0
! 8.6
! 8.7
! 8.9
! 9.0
|-
| Dot Product Unit Width in FP64 units (in bytes)
| colspan="1" {{yes| 4 (32)}}
| colspan="2" {{yes|tbd}}
|
| colspan="1" {{yes| 4 (32)}}
|-
| Dot Product Units per Tensor Core
| colspan="1" {{yes|4}}
| colspan="2" {{yes|tbd}}
|
| colspan="1" {{yes|8}}
|-
| Tensor Cores per SM partition
| colspan="5" {{yes|1}}
|-
| Full throughput (Bytes/cycle)<ref name="ReferenceC"/> per SM partition<ref name="product first 3 table rows"/>
| colspan="1" {{yes|128}}
| colspan="2" {{yes|tbd}}
|
| colspan="1" {{yes|256}}
|-
| Minimum cycles for warp-wide matrix calculation
| colspan="1" {{yes|16}}
| colspan="2" {{yes|tbd}}
|
|
|-
| Minimum Matrix Shape for full throughput (Bytes)<ref name="ReferenceD"/>
| colspan="3" {{yes|2048}}
|
|
|}

===Technical specification===
<div style="overflow-x:auto">
{| class="wikitable" style="font-size:85%;"
|- |-
! rowspan=2 | Technical specifications ! rowspan=2 | Technical specifications
! colspan=10 | Compute capability (version) ! colspan="23" | Compute capability (version)
|- |-
! 1.0 ! 1.0
Line 789: Line 906:
! 2.x ! 2.x
! 3.0 ! 3.0
! 3.2
! 3.5 ! 3.5
! 3.7 ! 3.7
! 5.0 ! 5.0
! 5.2 ! 5.2
! 5.3
! 6.0
! 6.1
! 6.2
! 7.0
! 7.2
! 7.5
! 8.0
! 8.6
! 8.7
! 8.9
! 9.0
|-
| Maximum number of resident grids per device<br />(concurrent kernel execution, can be lower for specific devices)
| colspan="4" {{yes|1}}
| colspan="2" {{yes|16}}
| colspan="1" {{yes|4}}
| colspan="4" {{yes|32}}
| colspan="1" {{yes|16}}
| colspan="1" {{yes|128}}
| colspan="1" {{yes|32}}
| colspan="1" {{yes|16}}
| colspan="1" {{yes|128}}
| colspan="1" {{yes|16}}
| colspan="6" {{yes|128}}
|- |-
! style="text-align:left;"| Maximum dimensionality of grid of thread blocks | Maximum dimensionality of grid of thread blocks
| colspan="4" {{yes|2}} | colspan="4" {{yes|2}}
| colspan="6" {{yes|3}} | colspan="19" {{yes|3}}
|- |-
! style="text-align:left;"| Maximum x-dimension of a grid of thread blocks | Maximum x-dimension of a grid of thread blocks
| colspan="5" {{yes|65535}} | colspan="5" {{yes|65535}}
| colspan="5" {{yes|2<sup>31</sup>-1}} | colspan="18" {{yes|2<sup>31</sup> − 1}}
|- |-
! style="text-align:left;"| Maximum y-, or z-dimension of a grid of thread blocks | Maximum y-, or z-dimension of a grid of thread blocks
| colspan="10" {{yes|65535}} | colspan="23" {{yes|65535}}
|- |-
! style="text-align:left;"| Maximum dimensionality of thread block | Maximum dimensionality of thread block
| colspan="10" {{yes|3}} | colspan="23" {{yes|3}}
|- |-
! style="text-align:left;"| Maximum x- or y-dimension of a block | Maximum x- or y-dimension of a block
| colspan="4" {{yes|512}} | colspan="4" {{yes|512}}
| colspan="6" {{yes|1024}} | colspan="19" {{yes|1024}}
|- |-
! style="text-align:left;"| Maximum z-dimension of a block | Maximum z-dimension of a block
| colspan="10" {{yes|64}} | colspan="23" {{yes|64}}
|- |-
! style="text-align:left;"| Maximum number of threads per block | Maximum number of threads per block
| colspan="4" {{yes|512}} | colspan="4" {{yes|512}}
| colspan="6" {{yes|1024}} | colspan="19" {{yes|1024}}
|- |-
! style="text-align:left;"| Warp size | Warp size
| colspan="10" {{yes|32}} | colspan="23" {{yes|32}}
|- |-
! style="text-align:left;"| Maximum number of resident blocks per multiprocessor | Maximum number of resident blocks per multiprocessor
| colspan="5" {{yes|8}} | colspan="5" {{yes|8}}
| colspan="3" {{yes|16}} | colspan="4" {{yes|16}}
| colspan="2" {{yes|32}} | colspan="8" {{yes|32}}
| colspan="1" {{yes|16}}
| colspan="1" {{yes|32}}
| colspan="2" {{yes|16}}
| colspan="1" {{yes|24}}
| colspan="1" {{yes|32}}
|- |-
! style="text-align:left;"| Maximum number of resident warps per multiprocessor | Maximum number of resident warps per multiprocessor
| colspan="2" {{yes|24}} | colspan="2" {{yes|24}}
| colspan="2" {{yes|32}} | colspan="2" {{yes|32}}
| {{yes|48}} | colspan="1" {{yes|48}}
| colspan="5" {{yes|64}} | colspan="12" {{yes|64}}
| colspan="1" {{yes|32}}
| colspan="1" {{yes|64}}
| colspan="3" {{yes|48}}
| colspan="1" {{yes|64}}
|- |-
! style="text-align:left;"| Maximum number of resident threads per multiprocessor | Maximum number of resident threads per multiprocessor
| colspan="2" {{yes|768}} | colspan="2" {{yes|768}}
| colspan="2" {{yes|1024}} | colspan="2" {{yes|1024}}
| {{yes|1536}} | colspan="1" {{yes|1536}}
| colspan="5" {{yes|2048}} | colspan="12" {{yes|2048}}
| colspan="1" {{yes|1024}}
| colspan="1" {{yes|2048}}
| colspan="3" {{yes|1536}}
| colspan="1" {{yes|2048}}
|- |-
! style="text-align:left;"| Number of 32-bit registers per multiprocessor | Number of 32-bit regular registers per multiprocessor
| colspan="2" {{yes|8 K}} | colspan="2" {{yes|8 K}}
| colspan="2" {{yes|16 K}} | colspan="2" {{yes|16 K}}
| {{yes|32 K}} | colspan="1" {{yes|32 K}}
| colspan="2" {{yes|64 K}} | colspan="3" {{yes|64 K}}
| colspan="1" {{yes|128 K}} | colspan="1" {{yes|128 K}}
| colspan="14" {{yes|64 K}}
|-
| Number of 32-bit uniform registers per multiprocessor
| colspan="17" {{no}}
| colspan="1" {{yes|2 K}}<ref name="ReferenceE">{{Cite arXiv| title=Dissecting the NVidia Turing T4 GPU via Microbenchmarking | eprint=1903.07486 | last1=Jia | first1=Zhe | last2=Maggioni | first2=Marco | last3=Smith | first3=Jeffrey | author4=Daniele Paolo Scarpazza | year=2019 | class=cs.DC }}</ref>
<ref>{{cite book | chapter-url=https://ieeexplore.ieee.org/document/8875651 | chapter=RTX ON – The NVIDIA TURING GPU | doi=10.1109/HOTCHIPS.2019.8875651 | title=2019 IEEE Hot Chips 31 Symposium (HCS) | year=2019 | last1=Burgess | first1=John | pages=1–27 | isbn=978-1-7281-2089-8 | s2cid=204822166 }}</ref>
| colspan="5" {{yes|}}
|-
| Maximum number of 32-bit registers per thread block
| colspan="2" {{yes|8 K}}
| colspan="2" {{yes|16 K}}
| colspan="1" {{yes|32 K}}
| colspan="1" {{yes|64 K}}
| colspan="1" {{yes|32 K}}
| colspan="4" {{yes|64 K}}
| colspan="1" {{yes|32 K}}
| colspan="2" {{yes|64 K}} | colspan="2" {{yes|64 K}}
| colspan="1" {{yes|32 K}}
| colspan="8" {{yes|64 K}}
|- |-
! style="text-align:left;"|Maximum number of 32-bit registers per thread | Maximum number of 32-bit regular registers per thread
| colspan="4" {{yes|128}} | colspan="4" {{yes|124}}
| colspan="2" {{yes|63}} | colspan="2" {{yes|63}}
| colspan="4" {{yes|255}} | colspan="17" {{yes|255}}
|- |-
! style="text-align:left;"| Maximum amount of shared memory per multiprocessor | Maximum number of 32-bit uniform registers per warp
| colspan="4" {{yes|16 KB}} | colspan="17" {{no}}
| colspan="3" {{yes|48 KB}} | colspan="1" {{yes|63}}<ref name="ReferenceE"/>
<ref>{{cite book | chapter-url=https://ieeexplore.ieee.org/document/8875651 | chapter=RTX ON – The NVIDIA TURING GPU | doi=10.1109/HOTCHIPS.2019.8875651 | title=2019 IEEE Hot Chips 31 Symposium (HCS) | year=2019 | last1=Burgess | first1=John | pages=1–27 | isbn=978-1-7281-2089-8 | s2cid=204822166 }}</ref>
| colspan="1" {{yes|112 KB}}
| colspan="1" {{yes|64 KB}} | colspan="5" {{yes|}}
| colspan="1" {{yes|96 KB}}
|- |-
| Amount of shared memory per multiprocessor<br/>(out of overall shared memory + L1 cache, where applicable)
! style="text-align:left;"| Number of shared memory banks
| colspan="4" {{yes|16 KiB}}
| colspan="1" {{yes|16 / 48 KiB (of 64 KiB)}}
| colspan="3" {{yes|16 / 32 / 48 KiB (of 64 KiB)}}
| colspan="1" {{yes|80 / 96 / 112 KiB (of 128 KiB)}}
| colspan="1" {{yes|64 KiB}}
| colspan="1" {{yes|96 KiB}}
| colspan="2" {{yes|64 KiB}}
| colspan="1" {{yes|96 KiB}}
| colspan="1" {{yes|64 KiB}}
| colspan="2" {{yes|0 / 8 / 16 / 32 / 64 / 96 KiB (of 128 KiB)}}
| colspan="1" {{yes|32 / 64 KiB (of 96 KiB)}}
| colspan="1" {{yes|0 / 8 / 16 / 32 / 64 / 100 / 132 / 164 KiB (of 192 KiB)}}
| colspan="1" {{yes|0 / 8 / 16 / 32 / 64 / 100 KiB (of 128 KiB)}}
| colspan="1" {{yes|0 / 8 / 16 / 32 / 64 / 100 / 132 / 164 KiB (of 192 KiB)}}
| colspan="1" {{yes|0 / 8 / 16 / 32 / 64 / 100 KiB (of 128 KiB)}}
| colspan="1" {{yes|0 / 8 / 16 / 32 / 64 / 100 / 132 / 164 / 196 / 228 KiB (of 256 KiB)}}
|-
| Maximum amount of shared memory per thread block
| colspan="4" {{yes|16 KiB}}
| colspan="11" {{yes|48 KiB}}
| colspan="1" {{yes|96 KiB}}
| colspan="1" {{yes|48 KiB}}
| colspan="1" {{yes|64 KiB}}
| colspan="1" {{yes|163 KiB}}
| colspan="1" {{yes|99 KiB}}
| colspan="1" {{yes|163 KiB}}
| colspan="1" {{yes|99 KiB}}
| colspan="1" {{yes|227 KiB}}
|-
| Number of shared memory banks
| colspan="4" {{yes|16}} | colspan="4" {{yes|16}}
| colspan="6" {{yes|32}} | colspan="19" {{yes|32}}
|- |-
! style="text-align:left;"| Amount of local memory per thread | Amount of local memory per thread
| colspan="4" {{yes|16 KB}} | colspan="4" {{yes|16 KiB}}
| colspan="6" {{yes|512 KB}} | colspan="19" {{yes|512 KiB}}
|- |-
| Constant memory size accessible by CUDA C/C++<br />(1 bank, PTX can access 11 banks, SASS can access 18 banks)
! style="text-align:left;"| Constant memory size
| colspan="10" {{yes|64 KB}} | colspan="23" {{yes|64 KiB}}
|- |-
! style="text-align:left;"| Cache working set per multiprocessor for constant memory | Cache working set per multiprocessor for constant memory
| colspan="8" {{yes|8 KB}} | colspan="12" {{yes|8 KiB}}
| colspan="2" {{yes|10 KB}} | colspan="1" {{yes|4 KiB}}
| colspan="10" {{yes|8 KiB}}
|- |-
! style="text-align:left;"| Cache working set per multiprocessor for texture memory | Cache working set per multiprocessor for texture memory
| colspan="4" {{yes| Device dependent, between 6&nbsp;KB and 8&nbsp;KB}} | colspan="3" {{yes| 16&nbsp;KiB per TPC}}
| colspan="2" {{yes| 12&nbsp;KB}} | colspan="1" {{yes| 24&nbsp;KiB per TPC}}
| colspan="2" {{yes| Between 12&nbsp;KB<br />and 48&nbsp;KB}} | colspan="1" {{yes| 12&nbsp;KiB}}
| colspan="2" {{yes| 24&nbsp;KB}} | colspan="4" {{yes| 12&nbsp;–&nbsp;48&nbsp;KiB}}<ref>dependent on device</ref>
| colspan="1" {{yes| 24&nbsp;KiB}}
| colspan="1" {{yes| 48&nbsp;KiB}}
| colspan="1" {{yes| 32&nbsp;KiB}}<ref name="Tegra X1">{{Cite web|url=https://developer.nvidia.com/content/tegra-x1|title=Tegra X1|date=9 January 2015 }}</ref>
| colspan="1" {{yes| 24&nbsp;KiB}}
| colspan="1" {{yes| 48&nbsp;KiB}}
| colspan="1" {{yes| 24&nbsp;KiB}}
| colspan="2" {{yes| 32&nbsp;–&nbsp;128&nbsp;KiB}}
| colspan="1" {{yes| 32&nbsp;–&nbsp;64&nbsp;KiB}}
| colspan="1" {{yes| 28&nbsp;–&nbsp;192&nbsp;KiB}}
| colspan="1" {{yes| 28&nbsp;–&nbsp;128&nbsp;KiB}}
| colspan="1" {{yes| 28&nbsp;–&nbsp;192&nbsp;KiB}}
| colspan="1" {{yes| 28&nbsp;–&nbsp;128&nbsp;KiB}}
| colspan="1" {{yes| 28&nbsp;–&nbsp;256&nbsp;KiB}}
|- |-
! style="text-align:left;"| Maximum width for 1D texture<br />reference bound to a CUDA array | Maximum width for 1D texture reference bound to a CUDA <br />array
| colspan="4" {{yes|8192}} | colspan="4" {{yes|8192}}
| colspan="6" {{yes|65536}} | colspan="8" {{yes|65536}}
| colspan="11" {{yes|131072}}
|- |-
! style="text-align:left;"| Maximum width for 1D texture<br />reference bound to linear memory | Maximum width for 1D texture reference bound to linear <br />memory
| colspan="10" {{yes| 2<sup>27</sup>}} | colspan="12" {{yes| 2<sup>27</sup>}}
| colspan="1" {{yes| 2<sup>28</sup>}}
| colspan="2" {{yes| 2<sup>27</sup>}}
| colspan="1" {{yes| 2<sup>28</sup>}}
| colspan="1" {{yes| 2<sup>27</sup>}}
| colspan="6" {{yes| 2<sup>28</sup>}}
|- |-
! style="text-align:left;"| Maximum width and number of layers<br />for a 1D layered texture reference | Maximum width and number of layers for a 1D layered <br />texture reference
| colspan="4" {{yes|8192 × 512}} | colspan="4" {{yes|8192 × 512}}
| colspan="6" {{yes|16384 × 2048}} | colspan="8" {{yes|16384 × 2048}}
| colspan="11" {{yes|32768 x 2048}}
|- |-
! style="text-align:left;"| Maximum width and height for 2D<br />texture reference bound to a CUDA array | Maximum width and height for 2D texture reference bound <br />to a CUDA array
| colspan="4" {{yes|65536 × 32768}} | colspan="4" {{yes|65536 × 32768}}
| colspan="6" {{yes|65536 × 65535}} | colspan="8" {{yes|65536 × 65535}}
| colspan="11" {{yes|131072 x 65536}}
|- |-
! style="text-align:left;"| Maximum width and height for 2D<br />texture reference bound to a linear memory | Maximum width and height for 2D texture reference bound <br />to a linear memory
| colspan="4" {{yes|65000 × 65000}} | colspan="9" {{yes|65000 x 65000}}
| colspan="6" {{yes|65000 × 65000}} | colspan="3" {{yes|65536 x 65536}}
| colspan="11" {{yes|131072 x 65000}}
|- |-
! style="text-align:left;"| Maximum width and height for 2D<br />texture reference bound to a CUDA array<br />supporting texture gather | Maximum width and height for 2D texture reference bound <br />to a CUDA array supporting texture gather
| colspan="4" {{no|N/A}} | colspan="4" {{n/a}}
| colspan="6" {{yes|16384 × 16384}} | colspan="8" {{yes|16384 x 16384}}
| colspan="11" {{yes|32768 x 32768}}
|- |-
! style="text-align:left;"| Maximum width, height, and number<br />of layers for a 2D layered texture reference | Maximum width, height, and number of layers for a 2D <br />layered texture reference
| colspan="4" {{yes|8192 × 8192 × 512}} | colspan="4" {{yes|8192 × 8192 × 512}}
| colspan="6" {{yes|16384 × 16384 × 2048}} | colspan="8" {{yes|16384 × 16384 × 2048}}
| colspan="11" {{yes|32768 x 32768 x 2048}}
|- |-
! style="text-align:left;"| Maximum width, height and depth<br />for a 3D texture reference bound to linear<br />memory or a CUDA array | Maximum width, height and depth for a 3D texture <br />reference bound to linear memory or a CUDA array
| colspan="5" {{yes|2048 × 2048 × 2048}} | colspan="5" {{yes|2048<sup>3</sup>}}
| colspan="5" {{yes|4096 × 4096 × 4096}} | colspan="7" {{yes|4096<sup>3</sup>}}
| colspan="11" {{yes|16384<sup>3</sup>}}
|- |-
! style="text-align:left;"| Maximum width (and height) for a cubemap<br />texture reference | Maximum width (and height) for a cubemap texture reference
| colspan="4" {{no|N/A}} | colspan="4" {{n/a}}
| colspan="6" {{yes|16384}} | colspan="8" {{yes|16384}}
| colspan="11" {{yes|32768}}
|- |-
! style="text-align:left;"| Maximum width (and height) and number<br />of layers for a cubemap layered texture reference | Maximum width (and height) and number of layers <br />for a cubemap layered texture reference
| colspan="4" {{no|N/A}} | colspan="4" {{n/a}}
| colspan="6" {{yes|16384 × 2046}} | colspan="8" {{yes|16384 × 2046}}
| colspan="11" {{yes|32768 × 2046}}
|- |-
! style="text-align:left;"| Maximum number of textures that<br />can be bound to a kernel | Maximum number of textures that can be bound to a <br />kernel
| colspan="5" {{yes|128}} | colspan="5" {{yes|128}}
| colspan="5" {{yes|256}} | colspan="18" {{yes|256}}
|- |-
! style="text-align:left;"| Maximum width for a 1D surface<br />reference bound to a CUDA array | Maximum width for a 1D surface reference bound to a <br />CUDA array
| colspan="4" rowspan="8" {{no|Not<br />supported}} | colspan="4" rowspan="8" {{no|Not<br />supported}}
| colspan="6" {{yes|65536}} | colspan="5" {{yes|65536}}
| colspan="3" {{yes|16384}}
| colspan="11" {{yes|32768}}
|- |-
! style="text-align:left;"| Maximum width and number of layers<br />for a 1D layered surface reference | Maximum width and number of layers for a 1D layered <br />surface reference
| colspan="6" {{yes|65536 × 2048}} | colspan="5" {{yes|65536 × 2048}}
| colspan="3" {{yes|16384 × 2048}}
| colspan="11" {{yes|32768 × 2048}}
|- |-
! style="text-align:left;"| Maximum width and height for a 2D<br />surface reference bound to a CUDA array | Maximum width and height for a 2D surface reference <br />bound to a CUDA array
| colspan="6" {{yes|65536 × 32768}} | colspan="5" {{yes|65536 × 32768}}
| colspan="3" {{yes|16384 × 65536}}
| colspan="11" {{yes|131072 × 65536}}
|- |-
! style="text-align:left;"| Maximum width, height, and number<br />of layers for a 2D layered surface reference | Maximum width, height, and number of layers for a 2D <br />layered surface reference
| colspan="6" {{yes|65536 × 32768 × 2048}} | colspan="5" {{yes|65536 × 32768 × 2048}}
| colspan="3" {{yes|16384 × 16384 × 2048}}
| colspan="11" {{yes|32768 × 32768 × 2048}}
|- |-
! style="text-align:left;"| Maximum width, height, and depth<br />for a 3D surface reference bound to a CUDA array | Maximum width, height, and depth for a 3D surface <br />reference bound to a CUDA array
| colspan="6" {{yes|65536 × 32768 × 2048}} | colspan="5" {{yes|65536 × 32768 × 2048}}
| colspan="3" {{yes|4096 × 4096 × 4096}}
| colspan="11" {{yes|16384 × 16384 × 16384}}
|- |-
! style="text-align:left;"| Maximum width (and height) for a cubemap<br />surface reference bound to a CUDA array | Maximum width (and height) for a cubemap surface reference bound to a CUDA array
| colspan="6" {{yes|32768}} | colspan="5" {{yes|32768}}
| colspan="3" {{yes|16384}}
| colspan="11" {{yes|32768}}
|- |-
! style="text-align:left;"| Maximum width (and height) and number<br />of layers for a cubemap layered surface reference | Maximum width and number of layers for a cubemap <br />layered surface reference
| colspan="6" {{yes|32768 × 2046}} | colspan="5" {{yes|32768 × 2046}}
| colspan="3" {{yes|16384 × 2046}}
| colspan="11" {{yes|32768 × 2046}}
|- |-
! style="text-align:left;"| Maximum number of surfaces that<br />can be bound to a kernel | Maximum number of surfaces that can be bound to a <br />kernel
| colspan="1" {{yes|8}} | colspan="1" {{yes|8}}
| colspan="5" {{yes|16}} | colspan="10" {{yes|16}}
| colspan="8" {{yes|32}}
|- |-
! style="text-align:left;"| Maximum number of instructions per<br />kernel | Maximum number of instructions per kernel
| colspan="4" {{yes|2 million}} | colspan="4" {{yes|2 million}}
| colspan="6" {{yes|512 million}} | colspan="19" {{yes|512 million}}
|- |-
| Maximum number of Thread Blocks per Thread Block Cluster<ref></ref>
! rowspan=2 | Technical specifications
| colspan="22" {{no}}
! 1.0
| {{yes|16}}
! 1.1
! 1.2
! 1.3
! 2.x
! 3.0
! 3.5
! 3.7
! 5.0
! 5.2
|- |-
! rowspan="2" |Technical specifications
! colspan=10 | Compute capability (version)
!1.0
!1.1
!1.2
!1.3
!2.x
!3.0
!3.2
!3.5
!3.7
!5.0
!5.2
!5.3
!6.0
!6.1
!6.2
!7.0
!7.2
!7.5
!8.0
!8.6
!8.7
!8.9
!9.0
|- |-
! colspan="23" |Compute capability (version)
|}
|}<ref></ref>
<ref></ref>
</div>


===Multiprocessor architecture===
{| class="wikitable" style="font-size: 85%; text-align: center; width: auto;"
<div style="overflow-x:auto">
{| class="wikitable" style="font-size:85%;"
|- |-
! rowspan= 2 | Architecture specifications ! rowspan= 2 | Architecture specifications
! colspan=11 | Compute capability (version) ! colspan=24 | Compute capability (version)
|- |-
! 1.0 ! 1.0
Line 981: Line 1,247:
! 2.1 ! 2.1
! 3.0 ! 3.0
! 3.2
! 3.5 ! 3.5
! 3.7 ! 3.7
! 5.0 ! 5.0
! 5.2 ! 5.2
! 5.3
! 6.0
! 6.1
! 6.2
! 7.0
! 7.2
! 7.5
! 8.0
! 8.6
! 8.7
! 8.9
! 9.0
|- |-
! style="text-align:left;"| Number of cores for integer and floating-point arithmetic functions operations | Number of ALU lanes for INT32 arithmetic operations
| rowspan="3" colspan="4" {{yes|8}}
| colspan="4" {{yes|8}}<ref>Cores perform only single-precision floating-point arithmetics. There is 1 double-precision floating-point unit.</ref>
| rowspan="3" colspan="1" {{yes|32}}
| rowspan="3" colspan="1" {{yes|48}}
| rowspan="3" colspan="4" {{yes|192<ref>can only execute 160 integer instructions according to programming guide</ref>}}
| rowspan="3" colspan="2" {{yes|128}}
| rowspan="4" colspan="1" {{yes|128}}
| rowspan="4" colspan="1" {{yes|64}}
| rowspan="3" colspan="1" {{yes|128}}
| rowspan="4" colspan="1" {{yes|128}}
| colspan="4" {{yes|64}}
| rowspan="2" colspan="2" {{yes|64}}
| colspan="2" {{yes|64}}
|-
| Number of ALU lanes for any INT32 or FP32 arithmetic operation
| colspan="4" {{n/a}}
| colspan="2" {{n/a}}
|-
| Number of ALU lanes for FP32 arithmetic operations
| rowspan="2" colspan="3" {{yes|64}}
| rowspan="1" colspan="3" {{yes|64}}
| rowspan="1" colspan="1" {{yes|128}}
| rowspan="2" colspan="1" {{yes|128}}
|-
| Number of ALU lanes for FP16x2 arithmetic operations
| colspan="12" {{no}}
| colspan="1" {{yes|1}}
| rowspan="1" colspan="1" {{yes|128<ref>128 according to . 64 from FP32 + 64 separate units?</ref>}}
| rowspan="1" colspan="2" {{yes|128<ref>64 by FP32 cores and 64 by flexible FP32/INT cores.</ref>}}
| rowspan="1" colspan="1" {{yes|64<ref>{{cite web|url=https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#arithmetic-instructions|title=CUDA C++ Programming Guide}}</ref>}}
|-
| Number of ALU lanes for FP64 arithmetic operations
| colspan="3" {{no}}
| {{yes|1}}
| colspan="1" {{yes|16 by FP32<ref>32 FP32 lanes combine to 16 FP64 lanes. Maybe lower depending on model.</ref>}}
| colspan="1" {{yes|4 by FP32<ref>only supported by 16 FP32 lanes, they combine to 4 FP64 lanes</ref>}}
| colspan="2" {{yes|8}}
| colspan="1" {{yes|8 / 64<ref name="depending on model">depending on model</ref>}}
| colspan="1" {{yes|64}}
| colspan="3" {{yes|4<ref>Effective speed, probably over FP32 ports. No description of actual FP64 cores.</ref>}}
| {{yes|32}}
| colspan="2" {{yes|4}}
| colspan="2" {{yes|32}}
| {{yes|2}}
| {{yes|32}}
| {{yes|2}}
| {{maybe|2?}}
| {{yes|2}}
| {{yes|64}}
|-
| Number of Load/Store Units
| colspan="1" {{yes|4 per 2 SM}}
| colspan="1" {{yes|8 per 2 SM}}
| colspan="1" {{yes|8 per 2 SM / 3 SM<ref name="depending on model">depending on model</ref>}}
| colspan="1" {{yes|8 per 3 SM}}
| colspan="2" {{yes|16}}
| colspan="7" {{yes|32}}
| colspan="1" {{yes|16}}
| colspan="5" {{yes|32}}
| colspan="4" {{yes|16}}
| colspan="1" {{yes|32}} | colspan="1" {{yes|32}}
| colspan="1" {{yes|48}}
| colspan="3" {{yes|192}}
| colspan="2" {{yes|128}}
|- |-
! style="text-align:left;"| Number of special function units for single-precision floating-point transcendental functions | Number of special function units for single-precision floating-point transcendental functions
| colspan="4" {{yes|2}} | colspan="4" {{yes|2<ref>Can also be used for integer additions and comparisons</ref>}}
| colspan="1" {{yes|4}} | colspan="1" {{yes|4}}
| colspan="1" {{yes|8}} | colspan="1" {{yes|8}}
| colspan="5" {{yes|32}} | colspan="6" {{yes|32}}
| colspan="2" {{yes|16}}
| colspan="2" {{yes|32}}
| colspan="8" {{yes|16}}
|- |-
! style="text-align:left;"| Number of texture filtering units for every texture address unit or ''render output unit'' (ROP) | Number of texture mapping units (TMU)
| colspan="4" {{yes|2}} | colspan="1" {{yes|4 per 2 SM}}
| colspan="1" {{yes|8 per 2 SM}}
| colspan="1" {{yes|8 per 2 / 3SM<ref name="depending on model">depending on model</ref>}}
| colspan="1" {{yes|8 per 3 SM}}
| colspan="1" {{yes|4}} | colspan="1" {{yes|4}}
| colspan="1" {{yes|4 / 8<ref name="depending on model">depending on model</ref>}}
| colspan="1" {{yes|16}}
| colspan="1" {{yes|8}} | colspan="1" {{yes|8}}
| colspan="3" {{yes|16}} | colspan="2" {{yes|16}}
| colspan="2" {{yes|8}} | colspan="6" {{yes|8}}
| colspan="8" {{yes|4}}
|- |-
| Number of ALU lanes for uniform INT32 arithmetic operations
! style="text-align:left;"| Number of warp schedulers
| colspan="18" {{no}}
| colspan="1" {{yes|2}}<ref>2 clock cycles/instruction for each SM partition {{cite book | chapter-url=https://ieeexplore.ieee.org/document/8875651 | chapter=RTX ON – The NVIDIA TURING GPU | doi=10.1109/HOTCHIPS.2019.8875651 | title=2019 IEEE Hot Chips 31 Symposium (HCS) | year=2019 | last1=Burgess | first1=John | pages=1–27 | isbn=978-1-7281-2089-8 | s2cid=204822166 }}</ref>
| colspan="5" {{yes|}}
|-
| Number of tensor cores
| colspan="16" {{no}}
| colspan="2" {{yes|8 (1st gen.)}}<ref name="inside-volta">{{cite web|url=https://devblogs.nvidia.com/inside-volta/|title=Inside Volta: The World's Most Advanced Data Center GPU|first1=Luke|last1=Durant|first2=Olivier|last2=Giroux|first3=Mark|last3=Harris|first4=Nick|last4=Stam|date=May 10, 2017|website=Nvidia developer blog}}</ref>
| {{yes|0 / 8<ref name="depending on model">depending on model</ref> (2nd gen.)}}
| colspan="3" {{yes|4 (3rd gen.)}}
| colspan="2" {{yes|4 (4th gen.)}}
|-
| Number of raytracing cores
| colspan="18" {{no}}
| {{yes|0 / 1<ref name="depending on model">depending on model</ref> (1st gen.)}}
| {{no}}
| {{yes|1 (2nd gen.)}}
| colspan="1" {{no}}
| {{yes|1 (3rd gen.)}}
| colspan="1" {{no}}
|-
| Number of SM Partitions = Processing Blocks<ref>The schedulers and dispatchers have dedicated execution units unlike with Fermi and Kepler.</ref>
| colspan="10" {{yes|1}}
| colspan="3" {{yes|4}}
| colspan="1" {{yes|2}}
| colspan="10" {{yes|4}}
|-
| Number of warp schedulers per SM partition
| colspan="4" {{yes|1}} | colspan="4" {{yes|1}}
| colspan="2" {{yes|2}} | colspan="2" {{yes|2}}
| colspan="5" {{yes|4}} | colspan="4" {{yes|4}}
| colspan="14" {{yes|1}}
|- |-
| Max number of new instructions issued each cycle by a single scheduler<ref>Dispatching can overlap concurrently, if it takes more than one cycle (when there are less execution units than 32/SM Partition)</ref>
! style="text-align:left;"| Number of instructions issued at once by scheduler
| colspan="5" {{yes|1}} | colspan="4" {{yes|2<ref>Can dual issue MAD pipe and SFU pipe</ref>}}
| colspan="1" {{yes|1}}
| colspan="6" {{yes|2}}<ref>No more than one scheduler can issue 2 instructions at once. The first scheduler is in charge of the warps with an odd ID and the second scheduler is in charge of the warps with an even ID.</ref>
| colspan="1" {{yes|2}}<ref>No more than one scheduler can issue 2 instructions at once. The first scheduler is in charge of warps with odd IDs. The second scheduler is in charge of warps with even IDs.</ref>
|}
| colspan="10" {{yes|2}}
| colspan="8" {{yes|1}}
|-
| Size of unified memory for data cache and shared memory
| colspan="3" {{yes|16 KiB<ref name="shared memory only, no data cache">shared memory only, no data cache</ref>}}
| colspan="1" {{yes|16 KiB<ref name="shared memory only, no data cache">shared memory only, no data cache</ref>}}
| colspan="5" {{yes|64 KiB}}
| colspan="1" {{yes|128 KiB}}
| colspan="1" {{yes|64 KiB SM + 24 KiB L1 (separate)<ref name="ReferenceA">shared memory separate, but L1 includes texture cache</ref>}}
| colspan="1" {{yes|96 KiB SM + 24 KiB L1 (separate)<ref name="ReferenceA">shared memory separate, but L1 includes texture cache</ref>}}
| colspan="1" {{yes|64 KiB SM + 24 KiB L1 (separate)<ref name="ReferenceA">shared memory separate, but L1 includes texture cache</ref>}}
| colspan="1" {{yes|64 KiB SM + 24 KiB L1 (separate)<ref name="ReferenceA">shared memory separate, but L1 includes texture cache</ref>}}
| colspan="1" {{yes|96 KiB SM + 24 KiB L1 (separate)<ref name="ReferenceA">shared memory separate, but L1 includes texture cache</ref>}}
| colspan="1" {{yes|64 KiB SM + 24 KiB L1 (separate)<ref name="ReferenceA">shared memory separate, but L1 includes texture cache</ref>}}
| colspan="2" {{yes|128 KiB}}
| {{yes|96 KiB}}<ref>{{Cite web|url=https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#architecture-7-x|title=<!--CUDA C Programming Guide-->H.6.1. Architecture|website=docs.nvidia.com|language=en-us|access-date=2019-05-13}}</ref>
| {{yes|192 KiB}}
| {{yes|128 KiB}}
| {{yes|192 KiB}}
| {{yes|128 KiB}}
| {{yes|256 KiB}}
|-
| Size of L3 instruction cache per GPU
| colspan="3"|
| {{yes|32 KiB}}<ref>{{Cite web|url=https://www.stuffedcow.net/files/gpuarch-ispass2010.pdf|title=Demystifying GPU Microarchitecture through Microbenchmarking}}</ref>
| colspan="2"|
| colspan="18" rowspan="2" {{maybe|use L2 Data Cache}}
|-
| Size of L2 instruction cache per Texture Processor Cluster (TPC)
| colspan="3"|
| {{yes|8 KiB}}
| colspan="2"|
|-
| Size of L1.5 instruction cache per SM<ref name="ReferenceF">{{Cite arXiv|title=Dissecting the NVIDIA Volta GPU Architecture via Microbenchmarking|eprint=1804.06826 |last1=Jia |first1=Zhe |last2=Maggioni |first2=Marco |last3=Staiger |first3=Benjamin |last4=Scarpazza |first4=Daniele P. |year=2018 |class=cs.DC }}</ref>
| rowspan="2" colspan="3"|
| rowspan="2" {{yes|4 KiB}}
| colspan="2"|
| colspan="3"|
| {{yes|32 KiB}}
|
| {{yes|32 KiB}}
| {{yes|48 KiB}}<ref name="Tegra X1"/>
| {{yes|128 KiB}}
| {{yes|32 KiB}}
|
| rowspan="2" {{yes|128 KiB}}
| rowspan="2" |
| rowspan="2" {{yes|~46 KiB}}<ref>{{Cite arXiv|title=Dissecting the NVidia Turing T4 GPU via Microbenchmarking|eprint=1903.07486 |last1=Jia |first1=Zhe |last2=Maggioni |first2=Marco |last3=Smith |first3=Jeffrey |author4=Daniele Paolo Scarpazza |year=2019 |class=cs.DC }}</ref>
| rowspan="2" {{yes|128 KiB}}<ref>{{Cite web|url=https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s33322/|title=Dissecting the Ampere GPU Architecture through Microbenchmarking}}</ref>
| rowspan="2" colspan="4"|
|-
| Size of L1 instruction cache per SM
| colspan="2"|
| colspan="3"|
| colspan="3" {{yes|8 KiB}}
|
| colspan=2 {{yes|8 KiB}}
|
|-
| Size of L0 instruction cache per SM partition
| colspan="10" {{maybe|only 1 partition per SM}}
| colspan="6" {{no}}
| {{yes|12 KiB}}
|
| {{yes|16 KiB?}}<ref>Note that {{Cite arXiv|title=Dissecting the NVidia Turing T4 GPU via Microbenchmarking|eprint=1903.07486 |last1=Jia |first1=Zhe |last2=Maggioni |first2=Marco |last3=Smith |first3=Jeffrey |author4=Daniele Paolo Scarpazza |year=2019 |class=cs.DC }} disagrees and states 2 KiB L0 instruction cache per SM partition and 16 KiB L1 instruction cache per SM</ref>
| {{yes|32 KiB}}
| colspan="4"|
|-
| Instruction Width<ref name="ReferenceF"/>
| colspan="6" {{yes|32 bits instructions and 64 bits instructions}}<ref>{{Cite web|url=https://github.com/hyqneuron/asfermi/Opcode|title=asfermi Opcode|website=] }}</ref>
| colspan="4" {{yes|64 bits instructions + 64 bits control logic every 7 instructions}}
| colspan="6" {{yes|64 bits instructions + 64 bits control logic every 3 instructions}}
| colspan="8" {{yes|128 bits combined instruction and control logic}}
|-
| Memory Bus Width per Memory Partition in bits
| colspan="12" {{yes|64 ((G)DDR)}}
| colspan="1" {{yes|32 ((G)DDR)}}
| {{yes|512 (HBM)}}
| colspan="2" {{yes|32 ((G)DDR)}}
| {{yes|512 (HBM)}}
| colspan="2" {{yes|32 ((G)DDR)}}
| {{yes|512 (HBM)}}
| colspan="3" {{yes|32 ((G)DDR)}}
| {{yes|512 (HBM)}}
|-
| L2 Cache per Memory Partition
| colspan="3" {{Yes|16 KiB<ref name="ReferenceB">for access with texture engine only</ref>}}
| colspan="1" {{Yes|32 KiB<ref name="ReferenceB">for access with texture engine only</ref>}}
| colspan="4" {{Yes|128 KiB}}
| colspan="2" {{Yes|256 KiB}}
| colspan="1" {{Yes|1 MiB}}
| colspan="1" {{Yes|512 KiB}}
| colspan="1" {{Yes|128 KiB}}
| colspan="1" {{Yes|512 KiB}}
| colspan="1" {{Yes|256 KiB}}
| colspan="1" {{Yes|128 KiB}}
| colspan="1" {{Yes|768 KiB}}
| colspan="1" {{Yes|64 KiB}}
| colspan="1" {{Yes|512 KiB}}
| colspan="1" {{Yes|4 MiB}}
| colspan="2" {{Yes|512 KiB}}
| colspan="1" {{Yes|8 MiB<ref>25% disabled on RTX 4090</ref>}}
| colspan="1" {{Yes|5 MiB}}
|-
| Number of Render Output Units (ROP) per memory partition (or per GPC in later models)
| colspan="4" {{Yes|4}}
| colspan="3" {{Yes|8}}
| colspan="1" {{Yes|4}}
| colspan="3" {{Yes|8}}
| colspan="1" {{Yes|16}}
| colspan="1" {{Yes|8}}
| colspan="1" {{Yes|12}}
| colspan="1" {{Yes|8}}
| colspan="1" {{Yes|4}}
| colspan="1" {{Yes|16}}
| colspan="1" {{Yes|2}}
| colspan="1" {{Yes|8}}
| colspan="1" {{Yes|16}}
| colspan="1" {{Yes|16 per GPC}}
| colspan="1" {{Yes|3 per GPC}}
| colspan="2" {{Yes|16 per GPC}}
|-
! rowspan= 2 | Architecture specifications
! 1.0
! 1.1
! 1.2
! 1.3
! 2.0
! 2.1
! 3.0
! 3.2
! 3.5
! 3.7
! 5.0
! 5.2
! 5.3
! 6.0
! 6.1
! 6.2
! 7.0
! 7.2
! 7.5
! 8.0
! 8.6
! 8.7
! 8.9
! 9.0
|-
! colspan=24 | Compute capability (version)
|}<ref>{{Cite web|url=https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#compute-capability-8-x|title=<!--CUDA C Programming Guide -->I.7. Compute Capability 8.x|website=docs.nvidia.com|language=en-us|access-date=2022-10-12}}</ref>
</div>
For more information read the Nvidia CUDA programming guide.<ref>{{cite web|url= http://developer.download.nvidia.com/compute/DevZone/docs/html/C/doc/CUDA_C_Programming_Guide.pdf |title=Appendix F. Features and Technical Specifications }}&nbsp;{{small|(3.2 MiB)}}, page 148 of 175 (Version 5.0 October 2012).</ref>


==Current and future usages of CUDA architecture==
For more information please visit this site: http://www.geeks3d.com/20100606/gpu-computing-nvidia-cuda-compute-capability-comparative-table/ and also read Nvidia CUDA programming guide.<ref>{{PDFlink||3.2 MiB}}, Page 148 of 175 (Version 5.0 October 2012)</ref>
* Accelerated rendering of 3D graphics
* Accelerated interconversion of video file formats
* Accelerated ], ] and ]
*], e.g. ] DNA sequencing BarraCUDA<ref>{{Cite web|url=https://www.biocentric.nl/biocentric/nvidia-cuda-bioinformatics-barracuda/|title=nVidia CUDA Bioinformatics: BarraCUDA|date=2019-07-19|website=BioCentric|language=en|access-date=2019-10-15}}</ref>
* Distributed calculations, such as predicting the native conformation of ]
* Medical analysis simulations, for example ] based on ] and ] scan images
* Physical simulations,<ref>{{Cite web|title=Part V: Physics Simulation|url=https://developer.nvidia.com/gpugems/gpugems3/part-v-physics-simulation|access-date=2020-09-11|website=NVIDIA Developer|language=en}}</ref> in particular in ]
* ] training in ] problems
* ] inference
* ]
* ] projects, such as ] and other projects using ] software
* ]
* Mining ]
* ] (SfM) software


== Comparison with competitors ==
==Example==
CUDA competes with other GPU computing stacks: ] and ].
This example code in ] loads a texture from an image into an array on the GPU:
<source lang="cpp">
texture<float, 2, cudaReadModeElementType> tex;


Whereas Nvidia's CUDA is closed-source, Intel's OneAPI and AMD's ROCm are open source.
void foo()
{
cudaArray* cu_array;


=== Intel OneAPI ===
// Allocate array
{{Main|OneAPI (compute acceleration)}}
cudaChannelFormatDesc description = cudaCreateChannelDesc<float>();
cudaMallocArray(&cu_array, &description, width, height);


'''oneAPI''' is an initiative based in open standards, created to support software development for multiple hardware architectures.<ref>{{Cite web |title=oneAPI Programming Model |url=https://www.oneapi.io/ |access-date=2024-07-27 |website=oneAPI.io |language=en-US}}</ref> The oneAPI libraries must implement open specifications that are discussed publicly by the Special Interest Groups, offering the possibility for any developer or organization to implement their own versions of oneAPI libraries.<ref>{{Cite web |title=Specifications {{!}} oneAPI |url=https://www.oneapi.io/spec/ |access-date=2024-07-27 |website=oneAPI.io |language=en-US}}</ref><ref>{{Cite web |title=oneAPI Specification — oneAPI Specification 1.3-rev-1 documentation |url=https://oneapi-spec.uxlfoundation.org/specifications/oneapi/v1.3-rev-1/ |access-date=2024-07-27 |website=oneapi-spec.uxlfoundation.org}}</ref>
// Copy image data to array
cudaMemcpyToArray(cu_array, image, width*height*sizeof(float), cudaMemcpyHostToDevice);


Originally made by Intel, other hardware adopters include Fujitsu and Huawei.
// Set texture parameters (default)
tex.addressMode = cudaAddressModeClamp;
tex.addressMode = cudaAddressModeClamp;
tex.filterMode = cudaFilterModePoint;
tex.normalized = false; // do not normalize coordinates


==== Unified Acceleration Foundation (UXL) ====
// Bind the array to the texture
cudaBindTextureToArray(tex, cu_array);


Unified Acceleration Foundation (UXL) is a new technology consortium working on the continuation of the OneAPI initiative, with the goal to create a new open standard accelerator software ecosystem, related open standards and specification projects through Working Groups and Special Interest Groups (SIGs). The goal is to offer open alternatives to Nvidia's CUDA. The main companies behind it are Intel, Google, ARM, Qualcomm, Samsung, Imagination, and VMware.<ref>{{Cite web |title=Exclusive: Behind the plot to break Nvidia's grip on AI by targeting software |website=] |url=https://www.reuters.com/technology/behind-plot-break-nvidias-grip-ai-by-targeting-software-2024-03-25/ |access-date=2024-04-05}}</ref>
// Run kernel
dim3 blockDim(16, 16, 1);
dim3 gridDim((width + blockDim.x - 1)/ blockDim.x, (height + blockDim.y - 1) / blockDim.y, 1);
kernel<<< gridDim, blockDim, 0 >>>(d_data, height, width);


=== AMD ROCm ===
// Unbind the array from the texture
{{Main|ROCm}}
cudaUnbindTexture(tex);
} //end foo()


'''ROCm'''<ref>{{Cite web|url=https://github.com/RadeonOpenCompute/ROCm/issues/1628|title=Question: What does ROCm stand for? · Issue #1628 · RadeonOpenCompute/ROCm|website=Github.com|access-date=January 18, 2022}}</ref> is an open source software stack for ] (GPU) programming from ] (AMD).
__global__ void kernel(float* odata, int height, int width)
{
unsigned int x = blockIdx.x*blockDim.x + threadIdx.x;
unsigned int y = blockIdx.y*blockDim.y + threadIdx.y;
if (x < width && y < height) {
float c = tex2D(tex, x, y);
odata = c;
}
}
</source>

Below is an example given in ] that computes the product of two arrays on the GPU. The unofficial Python language bindings can be obtained from ''PyCUDA''.<ref></ref>
<source lang="python">
import pycuda.compiler as comp
import pycuda.driver as drv
import numpy
import pycuda.autoinit

mod = comp.SourceModule("""
__global__ void multiply_them(float *dest, float *a, float *b)
{
const int i = threadIdx.x;
dest = a * b;
}
""")

multiply_them = mod.get_function("multiply_them")

a = numpy.random.randn(400).astype(numpy.float32)
b = numpy.random.randn(400).astype(numpy.float32)

dest = numpy.zeros_like(a)
multiply_them(
drv.Out(dest), drv.In(a), drv.In(b),
block=(400,1,1))

print dest-a*b

</source>

Additional Python bindings to simplify matrix multiplication operations can be found in the program ''pycublas''.<ref></ref>

<source lang="python">
import numpy
from pycublas import CUBLASMatrix
A = CUBLASMatrix( numpy.mat(],],numpy.float32) )
B = CUBLASMatrix( numpy.mat(],,],numpy.float32) )
C = A*B
print C.np_mat()
</source>

==Language bindings==
* ] -
* ] – ,
* F# -
* ] –
* ] –
* ] – , , , ,
* ] –
* ] –
* ] – Parallel Computing Toolbox, MATLAB Distributed Computing Server,<ref>{{cite web|title=MATLAB Adds GPGPU Support|url=http://www.hpcwire.com/features/MATLAB-Adds-GPGPU-Support-103307084.html|date=2010-09-20}}</ref> and 3rd party packages like ].
* ] – , , .NET kernel and host code, CURAND, CUBLAS, CUFFT
* ] – ,
* ] – Numba, NumbaPro, , , Theano
* ] –
* ] –

== Current and future usages of CUDA architecture ==
* Accelerated rendering of 3D graphics
* Accelerated interconversion of video file formats
* Accelerated ], ] and ]
* Distributed calculations, such as predicting the native conformation of ]
* Medical analysis simulations, for example ] based on ] and ] scan images.
* Physical simulations, in particular in ].
* ] training in ] problems
* ]
* ]
* Mining ]


==See also== ==See also==
* ] – an open standard from ] for programming a variety of platforms, including GPUs, with ''single-source'' modern C++, similar to higher-level CUDA '''Runtime''' API (''single-source'')
*] - A debugger for CUDA, OpenACC, and parallel applications
* ] – the Stanford University graphics group's compiler
*] - A standard for programming a variety of platforms, including GPUs
* ]
*] – the Stanford University graphics group's compiler
*] * ]
*] * ]
* ] – an API for computing on remote computers
*]
* ]
*] – An API for computing on remote computers
* ] – low-level, high-performance 3D graphics and computing API
*]
* ] – ray tracing API by NVIDIA
* ] (cubin) – a type of fat binary
* ] – by NEC for their vector processor


==References== ==References==
{{Reflist|30em}} {{Reflist|30em}}


== External links == == Further reading ==

* {{Official website|www.nvidia.com/object/cuda_home.html}}
* {{Cite journal |last1=Buck |first1=Ian |last2=Foley |first2=Tim |last3=Horn |first3=Daniel |last4=Sugerman |first4=Jeremy |last5=Fatahalian |first5=Kayvon |last6=Houston |first6=Mike |last7=Hanrahan |first7=Pat |date=2004-08-01 |title=Brook for GPUs: stream computing on graphics hardware |url=https://dl.acm.org/doi/10.1145/1015706.1015800 |journal=ACM Transactions on Graphics |volume=23 |issue=3 |pages=777–786 |doi=10.1145/1015706.1015800 |issn=0730-0301}}
* on ]
* {{Cite journal |last1=Nickolls |first1=John |last2=Buck |first2=Ian |last3=Garland |first3=Michael |last4=Skadron |first4=Kevin |date=2008-03-01 |title=Scalable Parallel Programming with CUDA: Is CUDA the parallel programming model that application developers have been waiting for? |url=https://dl.acm.org/doi/10.1145/1365490.1365500 |journal=Queue |volume=6 |issue=2 |pages=40–53 |doi=10.1145/1365490.1365500 |issn=1542-7730}}
*

==External links==
* {{Official website}}


{{Nvidia}} {{Nvidia}}
{{CPU technologies}} {{CPU technologies}}
{{Parallel computing}} {{Parallel computing}}

{{Authority control}}

__FORCETOC__


{{DEFAULTSORT:Cuda}} {{DEFAULTSORT:Cuda}}
Line 1,160: Line 1,615:
] ]
] ]
] ]
] ]

Latest revision as of 17:23, 11 January 2025

Parallel computing platform and programming model
This article has multiple issues. Please help improve it or discuss these issues on the talk page. (Learn how and when to remove these messages)
This article contains instructions, advice, or how-to content. Please help rewrite the content so that it is more encyclopedic or move it to Wikiversity, Wikibooks, or Wikivoyage. (February 2024)
This article contains a pro and con list. Please help rewriting it into consolidated sections based on topics. (February 2024)
(Learn how and when to remove this message)
CUDA
Developer(s)Nvidia
Initial releaseFebruary 15, 2007; 17 years ago (2007-02-15)
Stable release12.6 / August 2024; 5 months ago (2024-08)
Operating systemWindows, Linux
PlatformSupported GPUs
TypeGPGPU
LicenseProprietary
Websitedeveloper.nvidia.com/cuda-zone

In computing, CUDA is a proprietary parallel computing platform and application programming interface (API) that allows software to use certain types of graphics processing units (GPUs) for accelerated general-purpose processing, an approach called general-purpose computing on GPUs. CUDA was created by Nvidia in 2006. When it was first introduced, the name was an acronym for Compute Unified Device Architecture, but Nvidia later dropped the common use of the acronym and now rarely expands it.

CUDA is a software layer that gives direct access to the GPU's virtual instruction set and parallel computational elements for the execution of compute kernels. In addition to drivers and runtime kernels, the CUDA platform includes compilers, libraries and developer tools to help programmers accelerate their applications.

CUDA is designed to work with programming languages such as C, C++, Fortran and Python. This accessibility makes it easier for specialists in parallel programming to use GPU resources, in contrast to prior APIs like Direct3D and OpenGL, which require advanced skills in graphics programming. CUDA-powered GPUs also support programming frameworks such as OpenMP, OpenACC and OpenCL.

Background

Further information: Graphics processing unit

The graphics processing unit (GPU), as a specialized computer processor, addresses the demands of real-time high-resolution 3D graphics compute-intensive tasks. By 2012, GPUs had evolved into highly parallel multi-core systems allowing efficient manipulation of large blocks of data. This design is more effective than general-purpose central processing unit (CPUs) for algorithms in situations where processing large blocks of data is done in parallel, such as:

Ian Buck, while at Stanford in 2000, created an 8K gaming rig using 32 GeForce cards, then obtained a DARPA grant to perform general purpose parallel programming on GPUs. He then joined Nvidia, where since 2004 he has been overseeing CUDA development. In pushing for CUDA, Jensen Huang aimed for the Nvidia GPUs to become a general hardware for scientific computing. CUDA was released in 2007. Around 2015, the focus of CUDA changed to neural networks.

Ontology

The following table offers a non-exact description for the ontology of CUDA framework.

The ontology of CUDA framework
memory
(hardware) 		memory (code, or variable scoping) computation
(hardware) 		computation
(code syntax) 		computation
(code semantics)
RAM non-CUDA variables host program one routine call
VRAM,
GPU L2 cache 		global, const, texture device grid simultaneous call of the same subroutine on many processors
GPU L1 cache local, shared SM ("streaming multiprocessor") block individual subroutine call
warp = 32 threads SIMD instructions
GPU L0 cache,
register 		thread (aka. "SP", "streaming processor", "cuda core", but these names are now deprecated) analogous to individual scalar ops within a vector op

Programming abilities

Example of CUDA processing flow
  1. Copy data from main memory to GPU memory
  2. CPU initiates the GPU compute kernel
  3. GPU's CUDA cores execute the kernel in parallel
  4. Copy the resulting data from GPU memory to main memory

The CUDA platform is accessible to software developers through CUDA-accelerated libraries, compiler directives such as OpenACC, and extensions to industry-standard programming languages including C, C++, Fortran and Python. C/C++ programmers can use 'CUDA C/C++', compiled to PTX with nvcc, Nvidia's LLVM-based C/C++ compiler, or by clang itself. Fortran programmers can use 'CUDA Fortran', compiled with the PGI CUDA Fortran compiler from The Portland Group. Python programmers can use the cuNumeric library to accelerate applications on Nvidia GPUs.

In addition to libraries, compiler directives, CUDA C/C++ and CUDA Fortran, the CUDA platform supports other computational interfaces, including the Khronos Group's OpenCL, Microsoft's DirectCompute, OpenGL Compute Shader and C++ AMP. Third party wrappers are also available for Python, Perl, Fortran, Java, Ruby, Lua, Common Lisp, Haskell, R, MATLAB, IDL, Julia, and native support in Mathematica.

In the computer game industry, GPUs are used for graphics rendering, and for game physics calculations (physical effects such as debris, smoke, fire, fluids); examples include PhysX and Bullet. CUDA has also been used to accelerate non-graphical applications in computational biology, cryptography and other fields by an order of magnitude or more.

CUDA provides both a low level API (CUDA Driver API, non single-source) and a higher level API (CUDA Runtime API, single-source). The initial CUDA SDK was made public on 15 February 2007, for Microsoft Windows and Linux. Mac OS X support was later added in version 2.0, which supersedes the beta released February 14, 2008. CUDA works with all Nvidia GPUs from the G8x series onwards, including GeForce, Quadro and the Tesla line. CUDA is compatible with most standard operating systems.

CUDA 8.0 comes with the following libraries (for compilation & runtime, in alphabetical order):

  • cuBLAS – CUDA Basic Linear Algebra Subroutines library
  • CUDART – CUDA Runtime library
  • cuFFT – CUDA Fast Fourier Transform library
  • cuRAND – CUDA Random Number Generation library
  • cuSOLVER – CUDA based collection of dense and sparse direct solvers
  • cuSPARSE – CUDA Sparse Matrix library
  • NPP – NVIDIA Performance Primitives library
  • nvGRAPH – NVIDIA Graph Analytics library
  • NVML – NVIDIA Management Library
  • NVRTC – NVIDIA Runtime Compilation library for CUDA C++

CUDA 8.0 comes with these other software components:

  • nView – NVIDIA nView Desktop Management Software
  • NVWMI – NVIDIA Enterprise Management Toolkit
  • GameWorks PhysX – is a multi-platform game physics engine

CUDA 9.0–9.2 comes with these other components:

  • CUTLASS 1.0 – custom linear algebra algorithms,
  • NVIDIA Video Decoder was deprecated in CUDA 9.2; it is now available in NVIDIA Video Codec SDK

CUDA 10 comes with these other components:

  • nvJPEG – Hybrid (CPU and GPU) JPEG processing

CUDA 11.0–11.8 comes with these other components:

  • CUB is new one of more supported C++ libraries
  • MIG multi instance GPU support
  • nvJPEG2000 – JPEG 2000 encoder and decoder

Advantages

CUDA has several advantages over traditional general-purpose computation on GPUs (GPGPU) using graphics APIs:

  • Scattered reads – code can read from arbitrary addresses in memory.
  • Unified virtual memory (CUDA 4.0 and above)
  • Unified memory (CUDA 6.0 and above)
  • Shared memory – CUDA exposes a fast shared memory region that can be shared among threads. This can be used as a user-managed cache, enabling higher bandwidth than is possible using texture lookups.
  • Faster downloads and readbacks to and from the GPU
  • Full support for integer and bitwise operations, including integer texture lookups

Limitations

  • Whether for the host computer or the GPU device, all CUDA source code is now processed according to C++ syntax rules. This was not always the case. Earlier versions of CUDA were based on C syntax rules. As with the more general case of compiling C code with a C++ compiler, it is therefore possible that old C-style CUDA source code will either fail to compile or will not behave as originally intended.
  • Interoperability with rendering languages such as OpenGL is one-way, with OpenGL having access to registered CUDA memory but CUDA not having access to OpenGL memory.
  • Copying between host and device memory may incur a performance hit due to system bus bandwidth and latency (this can be partly alleviated with asynchronous memory transfers, handled by the GPU's DMA engine).
  • Threads should be running in groups of at least 32 for best performance, with total number of threads numbering in the thousands. Branches in the program code do not affect performance significantly, provided that each of 32 threads takes the same execution path; the SIMD execution model becomes a significant limitation for any inherently divergent task (e.g. traversing a space partitioning data structure during ray tracing).
  • No emulation or fallback functionality is available for modern revisions.
  • Valid C++ may sometimes be flagged and prevent compilation due to the way the compiler approaches optimization for target GPU device limitations.
  • C++ run-time type information (RTTI) and C++-style exception handling are only supported in host code, not in device code.
  • In single-precision on first generation CUDA compute capability 1.x devices, denormal numbers are unsupported and are instead flushed to zero, and the precision of both the division and square root operations are slightly lower than IEEE 754-compliant single precision math. Devices that support compute capability 2.0 and above support denormal numbers, and the division and square root operations are IEEE 754 compliant by default. However, users can obtain the prior faster gaming-grade math of compute capability 1.x devices if desired by setting compiler flags to disable accurate divisions and accurate square roots, and enable flushing denormal numbers to zero.
  • Unlike OpenCL, CUDA-enabled GPUs are only available from Nvidia as it is proprietary. Attempts to implement CUDA on other GPUs include:
    • Project Coriander: Converts CUDA C++11 source to OpenCL 1.2 C. A fork of CUDA-on-CL intended to run TensorFlow.
    • CU2CL: Convert CUDA 3.2 C++ to OpenCL C.
    • GPUOpen HIP: A thin abstraction layer on top of CUDA and ROCm intended for AMD and Nvidia GPUs. Has a conversion tool for importing CUDA C++ source. Supports CUDA 4.0 plus C++11 and float16.
    • ZLUDA is a drop-in replacement for CUDA on AMD GPUs and formerly Intel GPUs with near-native performance. The developer, Andrzej Janik, was separately contracted by both Intel and AMD to develop the software in 2021 and 2022, respectively. However, neither company decided to release it officially due to the lack of a business use case. AMD's contract included a clause that allowed Janik to release his code for AMD independently, allowing him to release the new version that only supports AMD GPUs.
    • chipStar can compile and run CUDA/HIP programs on advanced OpenCL 3.0 or Level Zero platforms.

Example

This example code in C++ loads a texture from an image into an array on the GPU: 

texture<float, 2, cudaReadModeElementType> tex;
void foo()
{
  cudaArray* cu_array;
  // Allocate array
  cudaChannelFormatDesc description = cudaCreateChannelDesc<float>();
  cudaMallocArray(&cu_array, &description, width, height);
  // Copy image data to array
  cudaMemcpyToArray(cu_array, image, width*height*sizeof(float), cudaMemcpyHostToDevice);
  // Set texture parameters (default)
  tex.addressMode = cudaAddressModeClamp;
  tex.addressMode = cudaAddressModeClamp;
  tex.filterMode = cudaFilterModePoint;
  tex.normalized = false; // do not normalize coordinates
  // Bind the array to the texture
  cudaBindTextureToArray(tex, cu_array);
  // Run kernel
  dim3 blockDim(16, 16, 1);
  dim3 gridDim((width + blockDim.x - 1)/ blockDim.x, (height + blockDim.y - 1) / blockDim.y, 1);
  kernel<<< gridDim, blockDim, 0 >>>(d_data, height, width);
  // Unbind the array from the texture
  cudaUnbindTexture(tex);
} //end foo()
__global__ void kernel(float* odata, int height, int width)
{
   unsigned int x = blockIdx.x*blockDim.x + threadIdx.x;
   unsigned int y = blockIdx.y*blockDim.y + threadIdx.y;
   if (x < width && y < height) {
      float c = tex2D(tex, x, y);
      odata = c;
   }
}

Below is an example given in Python that computes the product of two arrays on the GPU. The unofficial Python language bindings can be obtained from PyCUDA. 

import pycuda.compiler as comp
import pycuda.driver as drv
import numpy
import pycuda.autoinit
mod = comp.SourceModule(
    """
__global__ void multiply_them(float *dest, float *a, float *b)
{
  const int i = threadIdx.x;
  dest = a * b;
}
"""
)
multiply_them = mod.get_function("multiply_them")
a = numpy.random.randn(400).astype(numpy.float32)
b = numpy.random.randn(400).astype(numpy.float32)
dest = numpy.zeros_like(a)
multiply_them(drv.Out(dest), drv.In(a), drv.In(b), block=(400, 1, 1))
print(dest - a * b)

Additional Python bindings to simplify matrix multiplication operations can be found in the program pycublas. 

 
import numpy
from pycublas import CUBLASMatrix
A = CUBLASMatrix(numpy.mat(, ], numpy.float32))
B = CUBLASMatrix(numpy.mat(, , ], numpy.float32))
C = A * B
print(C.np_mat())

while CuPy directly replaces NumPy: 

import cupy
a = cupy.random.randn(400)
b = cupy.random.randn(400)
dest = cupy.zeros_like(a)
print(dest - a * b)

GPUs supported

Supported CUDA compute capability versions for CUDA SDK version and microarchitecture (by code name):

Compute capability (CUDA SDK support vs. microarchitecture)
CUDA SDK
version(s)		 Tesla Fermi Kepler
(early)		 Kepler
(late)		 Maxwell Pascal Volta Turing Ampere Ada
Lovelace		 Hopper Blackwell
1.0 1.0 – 1.1
1.1 1.0 – 1.1+x
2.0 1.0 – 1.1+x
2.1 – 2.3.1 1.0 – 1.3
3.0 – 3.1 1.0 2.0
3.2 1.0 2.1
4.0 – 4.2 1.0 2.1
5.0 – 5.5 1.0 3.5
6.0 1.0 3.2 3.5
6.5 1.1 3.7 5.x
7.0 – 7.5 2.0 5.x
8.0 2.0 6.x
9.0 – 9.2 3.0 7.0 – 7.2
10.0 – 10.2 3.0 7.5
11.0 3.5 8.0
11.1 – 11.4 3.5 8.6
11.5 – 11.7.1 3.5 8.7
11.8 3.5 8.9 9.0
12.0 – 12.5 5.0 9.0

Note: CUDA SDK 10.2 is the last official release for macOS, as support will not be available for macOS in newer releases.

CUDA compute capability by version with associated GPU semiconductors and GPU card models (separated by their various application areas):

Compute capability, GPU semiconductors and Nvidia GPU board products
Compute
capability
(version) 		Micro-
architecture 		GPUs GeForce Quadro, NVS Tesla/Datacenter Tegra,
Jetson,
DRIVE
1.0 Tesla G80 GeForce 8800 Ultra, GeForce 8800 GTX, GeForce 8800 GTS(G80) Quadro FX 5600, Quadro FX 4600, Quadro Plex 2100 S4 Tesla C870, Tesla D870, Tesla S870
1.1 G92, G94, G96, G98, G84, G86 GeForce GTS 250, GeForce 9800 GX2, GeForce 9800 GTX, GeForce 9800 GT, GeForce 8800 GTS(G92), GeForce 8800 GT, GeForce 9600 GT, GeForce 9500 GT, GeForce 9400 GT, GeForce 8600 GTS, GeForce 8600 GT, GeForce 8500 GT,
GeForce G110M, GeForce 9300M GS, GeForce 9200M GS, GeForce 9100M G, GeForce 8400M GT, GeForce G105M 		Quadro FX 4700 X2, Quadro FX 3700, Quadro FX 1800, Quadro FX 1700, Quadro FX 580, Quadro FX 570, Quadro FX 470, Quadro FX 380, Quadro FX 370, Quadro FX 370 Low Profile, Quadro NVS 450, Quadro NVS 420, Quadro NVS 290, Quadro NVS 295, Quadro Plex 2100 D4,
Quadro FX 3800M, Quadro FX 3700M, Quadro FX 3600M, Quadro FX 2800M, Quadro FX 2700M, Quadro FX 1700M, Quadro FX 1600M, Quadro FX 770M, Quadro FX 570M, Quadro FX 370M, Quadro FX 360M, Quadro NVS 320M, Quadro NVS 160M, Quadro NVS 150M, Quadro NVS 140M, Quadro NVS 135M, Quadro NVS 130M, Quadro NVS 450, Quadro NVS 420, Quadro NVS 295
1.2 GT218, GT216, GT215 GeForce GT 340*, GeForce GT 330*, GeForce GT 320*, GeForce 315*, GeForce 310*, GeForce GT 240, GeForce GT 220, GeForce 210,
GeForce GTS 360M, GeForce GTS 350M, GeForce GT 335M, GeForce GT 330M, GeForce GT 325M, GeForce GT 240M, GeForce G210M, GeForce 310M, GeForce 305M 		Quadro FX 380 Low Profile, Quadro FX 1800M, Quadro FX 880M, Quadro FX 380M,
Nvidia NVS 300, NVS 5100M, NVS 3100M, NVS 2100M, ION
1.3 GT200, GT200b GeForce GTX 295, GTX 285, GTX 280, GeForce GTX 275, GeForce GTX 260 Quadro FX 5800, Quadro FX 4800, Quadro FX 4800 for Mac, Quadro FX 3800, Quadro CX, Quadro Plex 2200 D2 Tesla C1060, Tesla S1070, Tesla M1060
2.0 Fermi GF100, GF110 GeForce GTX 590, GeForce GTX 580, GeForce GTX 570, GeForce GTX 480, GeForce GTX 470, GeForce GTX 465,
GeForce GTX 480M 		Quadro 6000, Quadro 5000, Quadro 4000, Quadro 4000 for Mac, Quadro Plex 7000,
Quadro 5010M, Quadro 5000M 		Tesla C2075, Tesla C2050/C2070, Tesla M2050/M2070/M2075/M2090
2.1 GF104, GF106 GF108, GF114, GF116, GF117, GF119 GeForce GTX 560 Ti, GeForce GTX 550 Ti, GeForce GTX 460, GeForce GTS 450, GeForce GTS 450*, GeForce GT 640 (GDDR3), GeForce GT 630, GeForce GT 620, GeForce GT 610, GeForce GT 520, GeForce GT 440, GeForce GT 440*, GeForce GT 430, GeForce GT 430*, GeForce GT 420*,
GeForce GTX 675M, GeForce GTX 670M, GeForce GT 635M, GeForce GT 630M, GeForce GT 625M, GeForce GT 720M, GeForce GT 620M, GeForce 710M, GeForce 610M, GeForce 820M, GeForce GTX 580M, GeForce GTX 570M, GeForce GTX 560M, GeForce GT 555M, GeForce GT 550M, GeForce GT 540M, GeForce GT 525M, GeForce GT 520MX, GeForce GT 520M, GeForce GTX 485M, GeForce GTX 470M, GeForce GTX 460M, GeForce GT 445M, GeForce GT 435M, GeForce GT 420M, GeForce GT 415M, GeForce 710M, GeForce 410M 		Quadro 2000, Quadro 2000D, Quadro 600,
Quadro 4000M, Quadro 3000M, Quadro 2000M, Quadro 1000M,
NVS 310, NVS 315, NVS 5400M, NVS 5200M, NVS 4200M
3.0 Kepler GK104, GK106, GK107 GeForce GTX 770, GeForce GTX 760, GeForce GT 740, GeForce GTX 690, GeForce GTX 680, GeForce GTX 670, GeForce GTX 660 Ti, GeForce GTX 660, GeForce GTX 650 Ti BOOST, GeForce GTX 650 Ti, GeForce GTX 650,
GeForce GTX 880M, GeForce GTX 870M, GeForce GTX 780M, GeForce GTX 770M, GeForce GTX 765M, GeForce GTX 760M, GeForce GTX 680MX, GeForce GTX 680M, GeForce GTX 675MX, GeForce GTX 670MX, GeForce GTX 660M, GeForce GT 750M, GeForce GT 650M, GeForce GT 745M, GeForce GT 645M, GeForce GT 740M, GeForce GT 730M, GeForce GT 640M, GeForce GT 640M LE, GeForce GT 735M, GeForce GT 730M 		Quadro K5000, Quadro K4200, Quadro K4000, Quadro K2000, Quadro K2000D, Quadro K600, Quadro K420,
Quadro K500M, Quadro K510M, Quadro K610M, Quadro K1000M, Quadro K2000M, Quadro K1100M, Quadro K2100M, Quadro K3000M, Quadro K3100M, Quadro K4000M, Quadro K5000M, Quadro K4100M, Quadro K5100M,
NVS 510, Quadro 410 		Tesla K10, GRID K340, GRID K520, GRID K2
3.2 GK20A Tegra K1,
Jetson TK1
3.5 GK110, GK208 GeForce GTX Titan Z, GeForce GTX Titan Black, GeForce GTX Titan, GeForce GTX 780 Ti, GeForce GTX 780, GeForce GT 640 (GDDR5), GeForce GT 630 v2, GeForce GT 730, GeForce GT 720, GeForce GT 710, GeForce GT 740M (64-bit, DDR3), GeForce GT 920M Quadro K6000, Quadro K5200 Tesla K40, Tesla K20x, Tesla K20
3.7 GK210 Tesla K80
5.0 Maxwell GM107, GM108 GeForce GTX 750 Ti, GeForce GTX 750, GeForce GTX 960M, GeForce GTX 950M, GeForce 940M, GeForce 930M, GeForce GTX 860M, GeForce GTX 850M, GeForce 845M, GeForce 840M, GeForce 830M Quadro K1200, Quadro K2200, Quadro K620, Quadro M2000M, Quadro M1000M, Quadro M600M, Quadro K620M, NVS 810 Tesla M10
5.2 GM200, GM204, GM206 GeForce GTX Titan X, GeForce GTX 980 Ti, GeForce GTX 980, GeForce GTX 970, GeForce GTX 960, GeForce GTX 950, GeForce GTX 750 SE,
GeForce GTX 980M, GeForce GTX 970M, GeForce GTX 965M 		Quadro M6000 24GB, Quadro M6000, Quadro M5000, Quadro M4000, Quadro M2000, Quadro M5500,
Quadro M5000M, Quadro M4000M, Quadro M3000M 		Tesla M4, Tesla M40, Tesla M6, Tesla M60
5.3 GM20B Tegra X1,
Jetson TX1,
Jetson Nano,
DRIVE CX,
DRIVE PX
6.0 Pascal GP100 Quadro GP100 Tesla P100
6.1 GP102, GP104, GP106, GP107, GP108 Nvidia TITAN Xp, Titan X,
GeForce GTX 1080 Ti, GTX 1080, GTX 1070 Ti, GTX 1070, GTX 1060,
GTX 1050 Ti, GTX 1050, GT 1030, GT 1010,
MX350, MX330, MX250, MX230, MX150, MX130, MX110 		Quadro P6000, Quadro P5000, Quadro P4000, Quadro P2200, Quadro P2000, Quadro P1000, Quadro P400, Quadro P500, Quadro P520, Quadro P600,
Quadro P5000 (mobile), Quadro P4000 (mobile), Quadro P3000 (mobile) 		Tesla P40, Tesla P6, Tesla P4
6.2 GP10B Tegra X2, Jetson TX2, DRIVE PX 2
7.0 Volta GV100 NVIDIA TITAN V Quadro GV100 Tesla V100, Tesla V100S
7.2 GV10B

GV11B

Tegra Xavier,
Jetson Xavier NX,
Jetson AGX Xavier,
DRIVE AGX Xavier,
DRIVE AGX Pegasus,
Clara AGX
7.5 Turing TU102, TU104, TU106, TU116, TU117 NVIDIA TITAN RTX,
GeForce RTX 2080 Ti, RTX 2080 Super, RTX 2080, RTX 2070 Super, RTX 2070, RTX 2060 Super, RTX 2060 12GB, RTX 2060,
GeForce GTX 1660 Ti, GTX 1660 Super, GTX 1660, GTX 1650 Super, GTX 1650, MX550, MX450 		Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 5000, Quadro RTX 4000, T1000, T600, T400
T1200 (mobile), T600 (mobile), T500 (mobile), Quadro T2000 (mobile), Quadro T1000 (mobile) 		Tesla T4
8.0 Ampere GA100 A100 80GB, A100 40GB, A30
8.6 GA102, GA103, GA104, GA106, GA107 GeForce RTX 3090 Ti, RTX 3090, RTX 3080 Ti, RTX 3080 12GB, RTX 3080, RTX 3070 Ti, RTX 3070, RTX 3060 Ti, RTX 3060, RTX 3050, RTX 3050 Ti (mobile), RTX 3050 (mobile), RTX 2050 (mobile), MX570 RTX A6000, RTX A5500, RTX A5000, RTX A4500, RTX A4000, RTX A2000
RTX A5000 (mobile), RTX A4000 (mobile), RTX A3000 (mobile), RTX A2000 (mobile) 		A40, A16, A10, A2
8.7 GA10B Jetson Orin Nano,
Jetson Orin NX,
Jetson AGX Orin,
DRIVE AGX Orin,
DRIVE AGX Pegasus OA,
Clara Holoscan
8.9 Ada Lovelace AD102, AD103, AD104, AD106, AD107 GeForce RTX 4090, RTX 4080 Super, RTX 4080, RTX 4070 Ti Super, RTX 4070 Ti, RTX 4070 Super, RTX 4070, RTX 4060 Ti, RTX 4060, RTX 4050 (mobile) RTX 6000 Ada, RTX 5880 Ada, RTX 5000 Ada, RTX 4500 Ada, RTX 4000 Ada, RTX 4000 SFF L40S, L40, L20, L4, L2
9.0 Hopper GH100 H200, H100
10.0 Blackwell GB100 B200, B100
10.x GB202, GB203, GB205, GB206, GB207 GeForce RTX 5090, RTX 5080, RTX 5070 Ti, RTX 5070 B40
Compute
capability
(version) 		Micro-
architecture 		GPUs GeForce Quadro, NVS Tesla/Datacenter Tegra,
Jetson,
DRIVE

'*' – OEM-only products

Version features and specifications

This section needs to be updated. The reason given is: Missing CUDA compute capability 10.x (Blackwell). Please help update this article to reflect recent events or newly available information. (March 2024)
Feature support (unlisted features are supported for all compute capabilities) Compute capability (version)
1.0, 1.1 1.2, 1.3 2.x 3.0 3.2 3.5, 3.7, 5.x, 6.x, 7.0, 7.2 7.5 8.x 9.0
Warp vote functions (__all(), __any()) No Yes
Warp vote functions (__ballot()) No Yes
Memory fence functions (__threadfence_system())
Synchronization functions (__syncthreads_count(), __syncthreads_and(), __syncthreads_or())
Surface functions
3D grid of thread blocks
Warp shuffle functions No Yes
Unified memory programming
Funnel shift No Yes
Dynamic parallelism No Yes
Uniform Datapath No Yes
Hardware-accelerated async-copy No Yes
Hardware-accelerated split arrive/wait barrier
Warp-level support for reduction ops
L2 cache residency management
DPX instructions for accelerated dynamic programming No Yes
Distributed shared memory
Thread block cluster
Tensor memory accelerator (TMA) unit
Feature support (unlisted features are supported for all compute capabilities) 1.0,1.1 1.2,1.3 2.x 3.0 3.2 3.5, 3.7, 5.x, 6.x, 7.0, 7.2 7.5 8.x 9.0
Compute capability (version)

Data types

Data type Operation Supported since
 Atomic Operation Supported since
for global memory 		Supported since
for shared memory
8-bit integer
signed/unsigned 		loading, storing, conversion 1.0 — —
16-bit integer
signed/unsigned 		general operations 1.0 atomicCAS() 3.5
32-bit integer
signed/unsigned 		general operations 1.0 atomic functions 1.1 1.2
64-bit integer
signed/unsigned 		general operations 1.0 atomic functions 1.2 2.0
any 128-bit trivially copyable type general operations No atomicExch, atomicCAS 9.0
16-bit floating point
FP16 		addition, subtraction,
multiplication, comparison,
warp shuffle functions, conversion 		5.3 half2 atomic addition 6.0
atomic addition 7.0
16-bit floating point
BF16 		addition, subtraction,
multiplication, comparison,
warp shuffle functions, conversion 		8.0 atomic addition 8.0
32-bit floating point general operations 1.0 atomicExch() 1.1 1.2
atomic addition 2.0
32-bit floating point float2 and float4 general operations No atomic addition 9.0
64-bit floating point general operations 1.3 atomic addition 6.0

Note: Any missing lines or empty entries do reflect some lack of information on that exact item.

Tensor cores

FMA per cycle per tensor core Supported since 7.0 7.2 7.5 Workstation 7.5 Desktop 8.0 8.6 Workstation 8.7 8.6 Desktop 8.9 Desktop 8.9 Workstation 9.0 10.0
Data Type For dense matrices For sparse matrices 1st Gen (8x/SM) 1st Gen? (8x/SM) 2nd Gen (8x/SM) 3rd Gen (4x/SM) 4th Gen (4x/SM) 5th Gen (4x/SM)
1-bit values (AND) 8.0 as
experimental 		No No 4096 2048 speed tbd
1-bit values (XOR) 7.5–8.9 as
experimental 		No 1024 Deprecated or removed?
4-bit integers 8.0–8.9 as
experimental 		256 1024 512
4-bit floating point FP4 (E2M1?) 10.0 No 4096
6-bit floating point FP6 (E3M2 and E2M3?) 10.0 No 2048
8-bit integers 7.2 8.0 No 128 128 512 256 1024 2048
8-bit floating point FP8 (E4M3 and E5M2) with FP16 accumulate 8.9 No 256
8-bit floating point FP8 (E4M3 and E5M2) with FP32 accumulate
16-bit floating point FP16 with FP16 accumulate 7.0 8.0 64 64 64 256 128 512 1024
16-bit floating point FP16 with FP32 accumulate 32 64 128
16-bit floating point BF16 with FP32 accumulate 7.5 8.0 No 64
32-bit (19 bits used) floating point TF32 speed tbd (32?) 128 32 64 256 512
64-bit floating point 8.0 No No 16 speed tbd 32 16

Note: Any missing lines or empty entries do reflect some lack of information on that exact item.

Tensor Core Composition 7.0 7.2, 7.5 8.0, 8.6 8.7 8.9 9.0
Dot Product Unit Width in FP16 units (in bytes) 4 (8) 8 (16) 4 (8) 16 (32)
Dot Product Units per Tensor Core 16 32
Tensor Cores per SM partition 2 1
Full throughput (Bytes/cycle) per SM partition 256 512 256 1024
FP Tensor Cores: Minimum cycles for warp-wide matrix calculation 8 4 8
FP Tensor Cores: Minimum Matrix Shape for full throughput (Bytes) 2048
INT Tensor Cores: Minimum cycles for warp-wide matrix calculation No 4
INT Tensor Cores: Minimum Matrix Shape for full throughput (Bytes) No 1024 2048 1024

FP64 Tensor Core Composition 8.0 8.6 8.7 8.9 9.0
Dot Product Unit Width in FP64 units (in bytes) 4 (32) tbd 4 (32)
Dot Product Units per Tensor Core 4 tbd 8
Tensor Cores per SM partition 1
Full throughput (Bytes/cycle) per SM partition 128 tbd 256
Minimum cycles for warp-wide matrix calculation 16 tbd
Minimum Matrix Shape for full throughput (Bytes) 2048

Technical specification

Technical specifications Compute capability (version)
1.0 1.1 1.2 1.3 2.x 3.0 3.2 3.5 3.7 5.0 5.2 5.3 6.0 6.1 6.2 7.0 7.2 7.5 8.0 8.6 8.7 8.9 9.0
Maximum number of resident grids per device
(concurrent kernel execution, can be lower for specific devices) 		1 16 4 32 16 128 32 16 128 16 128
Maximum dimensionality of grid of thread blocks 2 3
Maximum x-dimension of a grid of thread blocks 65535 2 − 1
Maximum y-, or z-dimension of a grid of thread blocks 65535
Maximum dimensionality of thread block 3
Maximum x- or y-dimension of a block 512 1024
Maximum z-dimension of a block 64
Maximum number of threads per block 512 1024
Warp size 32
Maximum number of resident blocks per multiprocessor 8 16 32 16 32 16 24 32
Maximum number of resident warps per multiprocessor 24 32 48 64 32 64 48 64
Maximum number of resident threads per multiprocessor 768 1024 1536 2048 1024 2048 1536 2048
Number of 32-bit regular registers per multiprocessor 8 K 16 K 32 K 64 K 128 K 64 K
Number of 32-bit uniform registers per multiprocessor No 2 K

Maximum number of 32-bit registers per thread block 8 K 16 K 32 K 64 K 32 K 64 K 32 K 64 K 32 K 64 K
Maximum number of 32-bit regular registers per thread 124 63 255
Maximum number of 32-bit uniform registers per warp No 63

Amount of shared memory per multiprocessor
(out of overall shared memory + L1 cache, where applicable) 		16 KiB 16 / 48 KiB (of 64 KiB) 16 / 32 / 48 KiB (of 64 KiB) 80 / 96 / 112 KiB (of 128 KiB) 64 KiB 96 KiB 64 KiB 96 KiB 64 KiB 0 / 8 / 16 / 32 / 64 / 96 KiB (of 128 KiB) 32 / 64 KiB (of 96 KiB) 0 / 8 / 16 / 32 / 64 / 100 / 132 / 164 KiB (of 192 KiB) 0 / 8 / 16 / 32 / 64 / 100 KiB (of 128 KiB) 0 / 8 / 16 / 32 / 64 / 100 / 132 / 164 KiB (of 192 KiB) 0 / 8 / 16 / 32 / 64 / 100 KiB (of 128 KiB) 0 / 8 / 16 / 32 / 64 / 100 / 132 / 164 / 196 / 228 KiB (of 256 KiB)
Maximum amount of shared memory per thread block 16 KiB 48 KiB 96 KiB 48 KiB 64 KiB 163 KiB 99 KiB 163 KiB 99 KiB 227 KiB
Number of shared memory banks 16 32
Amount of local memory per thread 16 KiB 512 KiB
Constant memory size accessible by CUDA C/C++
(1 bank, PTX can access 11 banks, SASS can access 18 banks) 		64 KiB
Cache working set per multiprocessor for constant memory 8 KiB 4 KiB 8 KiB
Cache working set per multiprocessor for texture memory 16 KiB per TPC 24 KiB per TPC 12 KiB 12 – 48 KiB 24 KiB 48 KiB 32 KiB 24 KiB 48 KiB 24 KiB 32 – 128 KiB 32 – 64 KiB 28 – 192 KiB 28 – 128 KiB 28 – 192 KiB 28 – 128 KiB 28 – 256 KiB
Maximum width for 1D texture reference bound to a CUDA
array 		8192 65536 131072
Maximum width for 1D texture reference bound to linear
memory 		2 2 2 2 2 2
Maximum width and number of layers for a 1D layered
texture reference 		8192 × 512 16384 × 2048 32768 x 2048
Maximum width and height for 2D texture reference bound
to a CUDA array 		65536 × 32768 65536 × 65535 131072 x 65536
Maximum width and height for 2D texture reference bound
to a linear memory 		65000 x 65000 65536 x 65536 131072 x 65000
Maximum width and height for 2D texture reference bound
to a CUDA array supporting texture gather 		— 16384 x 16384 32768 x 32768
Maximum width, height, and number of layers for a 2D
layered texture reference 		8192 × 8192 × 512 16384 × 16384 × 2048 32768 x 32768 x 2048
Maximum width, height and depth for a 3D texture
reference bound to linear memory or a CUDA array 		2048 4096 16384
Maximum width (and height) for a cubemap texture reference — 16384 32768
Maximum width (and height) and number of layers
for a cubemap layered texture reference 		— 16384 × 2046 32768 × 2046
Maximum number of textures that can be bound to a
kernel 		128 256
Maximum width for a 1D surface reference bound to a
CUDA array 		Not
supported 		65536 16384 32768
Maximum width and number of layers for a 1D layered
surface reference 		65536 × 2048 16384 × 2048 32768 × 2048
Maximum width and height for a 2D surface reference
bound to a CUDA array 		65536 × 32768 16384 × 65536 131072 × 65536
Maximum width, height, and number of layers for a 2D
layered surface reference 		65536 × 32768 × 2048 16384 × 16384 × 2048 32768 × 32768 × 2048
Maximum width, height, and depth for a 3D surface
reference bound to a CUDA array 		65536 × 32768 × 2048 4096 × 4096 × 4096 16384 × 16384 × 16384
Maximum width (and height) for a cubemap surface reference bound to a CUDA array 32768 16384 32768
Maximum width and number of layers for a cubemap
layered surface reference 		32768 × 2046 16384 × 2046 32768 × 2046
Maximum number of surfaces that can be bound to a
kernel 		8 16 32
Maximum number of instructions per kernel 2 million 512 million
Maximum number of Thread Blocks per Thread Block Cluster No 16
Technical specifications 1.0 1.1 1.2 1.3 2.x 3.0 3.2 3.5 3.7 5.0 5.2 5.3 6.0 6.1 6.2 7.0 7.2 7.5 8.0 8.6 8.7 8.9 9.0
Compute capability (version)

Multiprocessor architecture

Architecture specifications Compute capability (version)
1.0 1.1 1.2 1.3 2.0 2.1 3.0 3.2 3.5 3.7 5.0 5.2 5.3 6.0 6.1 6.2 7.0 7.2 7.5 8.0 8.6 8.7 8.9 9.0
Number of ALU lanes for INT32 arithmetic operations 8 32 48 192 128 128 64 128 128 64 64 64
Number of ALU lanes for any INT32 or FP32 arithmetic operation — —
Number of ALU lanes for FP32 arithmetic operations 64 64 128 128
Number of ALU lanes for FP16x2 arithmetic operations No 1 128 128 64
Number of ALU lanes for FP64 arithmetic operations No 1 16 by FP32 4 by FP32 8 8 / 64 64 4 32 4 32 2 32 2 2? 2 64
Number of Load/Store Units 4 per 2 SM 8 per 2 SM 8 per 2 SM / 3 SM 8 per 3 SM 16 32 16 32 16 32
Number of special function units for single-precision floating-point transcendental functions 2 4 8 32 16 32 16
Number of texture mapping units (TMU) 4 per 2 SM 8 per 2 SM 8 per 2 / 3SM 8 per 3 SM 4 4 / 8 16 8 16 8 4
Number of ALU lanes for uniform INT32 arithmetic operations No 2
Number of tensor cores No 8 (1st gen.) 0 / 8 (2nd gen.) 4 (3rd gen.) 4 (4th gen.)
Number of raytracing cores No 0 / 1 (1st gen.) No 1 (2nd gen.) No 1 (3rd gen.) No
Number of SM Partitions = Processing Blocks 1 4 2 4
Number of warp schedulers per SM partition 1 2 4 1
Max number of new instructions issued each cycle by a single scheduler 2 1 2 2 1
Size of unified memory for data cache and shared memory 16 KiB 16 KiB 64 KiB 128 KiB 64 KiB SM + 24 KiB L1 (separate) 96 KiB SM + 24 KiB L1 (separate) 64 KiB SM + 24 KiB L1 (separate) 64 KiB SM + 24 KiB L1 (separate) 96 KiB SM + 24 KiB L1 (separate) 64 KiB SM + 24 KiB L1 (separate) 128 KiB 96 KiB 192 KiB 128 KiB 192 KiB 128 KiB 256 KiB
Size of L3 instruction cache per GPU 32 KiB use L2 Data Cache
Size of L2 instruction cache per Texture Processor Cluster (TPC) 8 KiB
Size of L1.5 instruction cache per SM 4 KiB 32 KiB 32 KiB 48 KiB 128 KiB 32 KiB 128 KiB ~46 KiB 128 KiB
Size of L1 instruction cache per SM 8 KiB 8 KiB
Size of L0 instruction cache per SM partition only 1 partition per SM No 12 KiB 16 KiB? 32 KiB
Instruction Width 32 bits instructions and 64 bits instructions 64 bits instructions + 64 bits control logic every 7 instructions 64 bits instructions + 64 bits control logic every 3 instructions 128 bits combined instruction and control logic
Memory Bus Width per Memory Partition in bits 64 ((G)DDR) 32 ((G)DDR) 512 (HBM) 32 ((G)DDR) 512 (HBM) 32 ((G)DDR) 512 (HBM) 32 ((G)DDR) 512 (HBM)
L2 Cache per Memory Partition 16 KiB 32 KiB 128 KiB 256 KiB 1 MiB 512 KiB 128 KiB 512 KiB 256 KiB 128 KiB 768 KiB 64 KiB 512 KiB 4 MiB 512 KiB 8 MiB 5 MiB
Number of Render Output Units (ROP) per memory partition (or per GPC in later models) 4 8 4 8 16 8 12 8 4 16 2 8 16 16 per GPC 3 per GPC 16 per GPC
Architecture specifications 1.0 1.1 1.2 1.3 2.0 2.1 3.0 3.2 3.5 3.7 5.0 5.2 5.3 6.0 6.1 6.2 7.0 7.2 7.5 8.0 8.6 8.7 8.9 9.0
Compute capability (version)

For more information read the Nvidia CUDA programming guide.

Current and future usages of CUDA architecture

Comparison with competitors

CUDA competes with other GPU computing stacks: Intel OneAPI and AMD ROCm.

Whereas Nvidia's CUDA is closed-source, Intel's OneAPI and AMD's ROCm are open source.

Intel OneAPI

Main article: OneAPI (compute acceleration)

oneAPI is an initiative based in open standards, created to support software development for multiple hardware architectures. The oneAPI libraries must implement open specifications that are discussed publicly by the Special Interest Groups, offering the possibility for any developer or organization to implement their own versions of oneAPI libraries.

Originally made by Intel, other hardware adopters include Fujitsu and Huawei.

Unified Acceleration Foundation (UXL)

Unified Acceleration Foundation (UXL) is a new technology consortium working on the continuation of the OneAPI initiative, with the goal to create a new open standard accelerator software ecosystem, related open standards and specification projects through Working Groups and Special Interest Groups (SIGs). The goal is to offer open alternatives to Nvidia's CUDA. The main companies behind it are Intel, Google, ARM, Qualcomm, Samsung, Imagination, and VMware.

AMD ROCm

Main article: ROCm

ROCm is an open source software stack for graphics processing unit (GPU) programming from Advanced Micro Devices (AMD).

See also

References

  1. ^ Shah, Agam. "Nvidia not totally against third parties making CUDA chips". www.theregister.com. Retrieved 2024-04-25.
  2. "Nvidia CUDA Home Page". 18 July 2017.
  3. Shimpi, Anand Lal; Wilson, Derek (November 8, 2006). "Nvidia's GeForce 8800 (G80): GPUs Re-architected for DirectX 10". AnandTech. Retrieved May 16, 2015.
  4. "Introduction — nsight-visual-studio-edition 12.6 documentation". docs.nvidia.com. Retrieved 2024-10-10.
  5. ^ Abi-Chahla, Fedy (June 18, 2008). "Nvidia's CUDA: The End of the CPU?". Tom's Hardware. Retrieved May 17, 2015.
  6. Zunitch, Peter (2018-01-24). "CUDA vs. OpenCL vs. OpenGL". Videomaker. Retrieved 2018-09-16.
  7. "OpenCL". NVIDIA Developer. 2013-04-24. Retrieved 2019-11-04.
  8. Witt, Stephen (2023-11-27). "How Jensen Huang's Nvidia Is Powering the A.I. Revolution". The New Yorker. ISSN 0028-792X. Retrieved 2023-12-10.
  9. "CUDA LLVM Compiler". 7 May 2012.
  10. First OpenCL demo on a GPU on YouTube
  11. DirectCompute Ocean Demo Running on Nvidia CUDA-enabled GPU on YouTube
  12. Vasiliadis, Giorgos; Antonatos, Spiros; Polychronakis, Michalis; Markatos, Evangelos P.; Ioannidis, Sotiris (September 2008). "Gnort: High Performance Network Intrusion Detection Using Graphics Processors" (PDF). Recent Advances in Intrusion Detection. Lecture Notes in Computer Science. Vol. 5230. pp. 116–134. doi:10.1007/978-3-540-87403-4_7. ISBN 978-3-540-87402-7.
  13. Schatz, Michael C.; Trapnell, Cole; Delcher, Arthur L.; Varshney, Amitabh (2007). "High-throughput sequence alignment using Graphics Processing Units". BMC Bioinformatics. 8: 474. doi:10.1186/1471-2105-8-474. PMC 2222658. PMID 18070356.
  14. Manavski, Svetlin A.; Giorgio, Valle (2008). "CUDA compatible GPU cards as efficient hardware accelerators for Smith-Waterman sequence alignment". BMC Bioinformatics. 10 (Suppl 2): S10. doi:10.1186/1471-2105-9-S2-S10. PMC 2323659. PMID 18387198.
  15. "Pyrit – Google Code".
  16. "Use your Nvidia GPU for scientific computing". BOINC. 2008-12-18. Archived from the original on 2008-12-28. Retrieved 2017-08-08.
  17. "Nvidia CUDA Software Development Kit (CUDA SDK) – Release Notes Version 2.0 for MAC OS X". Archived from the original on 2009-01-06.
  18. "CUDA 1.1 – Now on Mac OS X". February 14, 2008. Archived from the original on November 22, 2008.
  19. "CUDA 11 Features Revealed". 14 May 2020.
  20. "CUDA Toolkit 11.1 Introduces Support for GeForce RTX 30 Series and Quadro RTX Series GPUs". 23 September 2020.
  21. "Enhancing Memory Allocation with New NVIDIA CUDA 11.2 Features". 16 December 2020.
  22. "Exploring the New Features of CUDA 11.3". 16 April 2021.
  23. Silberstein, Mark; Schuster, Assaf; Geiger, Dan; Patney, Anjul; Owens, John D. (2008). "Efficient computation of sum-products on GPUs through software-managed cache" (PDF). Proceedings of the 22nd annual international conference on Supercomputing – ICS '08 (PDF). Proceedings of the 22nd annual international conference on Supercomputing – ICS '08. pp. 309–318. doi:10.1145/1375527.1375572. ISBN 978-1-60558-158-3.
  24. "CUDA C Programming Guide v8.0" (PDF). nVidia Developer Zone. January 2017. p. 19. Retrieved 22 March 2017.
  25. "NVCC forces c++ compilation of .cu files". 29 November 2011.
  26. Whitehead, Nathan; Fit-Florea, Alex. "Precision & Performance: Floating Point and IEEE 754 Compliance for Nvidia GPUs" (PDF). Nvidia. Retrieved November 18, 2014.
  27. "CUDA-Enabled Products". CUDA Zone. Nvidia Corporation. Retrieved 2008-11-03.
  28. "Coriander Project: Compile CUDA Codes To OpenCL, Run Everywhere". Phoronix.
  29. Perkins, Hugh (2017). "cuda-on-cl" (PDF). IWOCL. Retrieved August 8, 2017.
  30. "hughperkins/coriander: Build NVIDIA® CUDA™ code for OpenCL™ 1.2 devices". GitHub. May 6, 2019.
  31. "CU2CL Documentation". chrec.cs.vt.edu.
  32. "GitHub – vosen/ZLUDA". GitHub.
  33. Larabel, Michael (2024-02-12), "AMD Quietly Funded A Drop-In CUDA Implementation Built On ROCm: It's Now Open-Source", Phoronix, retrieved 2024-02-12
  34. "GitHub – chip-spv/chipStar". GitHub.
  35. "PyCUDA".
  36. "pycublas". Archived from the original on 2009-04-20. Retrieved 2017-08-08.
  37. "CuPy". Retrieved 2020-01-08.
  38. "NVIDIA CUDA Programming Guide. Version 1.0" (PDF). June 23, 2007.
  39. "NVIDIA CUDA Programming Guide. Version 2.1" (PDF). December 8, 2008.
  40. "NVIDIA CUDA Programming Guide. Version 2.2" (PDF). April 2, 2009.
  41. "NVIDIA CUDA Programming Guide. Version 2.2.1" (PDF). May 26, 2009.
  42. "NVIDIA CUDA Programming Guide. Version 2.3.1" (PDF). August 26, 2009.
  43. "NVIDIA CUDA Programming Guide. Version 3.0" (PDF). February 20, 2010.
  44. "NVIDIA CUDA C Programming Guide. Version 3.1.1" (PDF). July 21, 2010.
  45. "NVIDIA CUDA C Programming Guide. Version 3.2" (PDF). November 9, 2010.
  46. "CUDA 11.0 Release Notes". NVIDIA Developer.
  47. "CUDA 11.1 Release Notes". NVIDIA Developer.
  48. "CUDA 11.5 Release Notes". NVIDIA Developer.
  49. "CUDA 11.8 Release Notes". NVIDIA Developer.
  50. "NVIDIA Quadro NVS 420 Specs". TechPowerUp GPU Database. 25 August 2023.
  51. Larabel, Michael (March 29, 2017). "NVIDIA Rolls Out Tegra X2 GPU Support In Nouveau". Phoronix. Retrieved August 8, 2017.
  52. Nvidia Xavier Specs on TechPowerUp (preliminary)
  53. "Welcome — Jetson LinuxDeveloper Guide 34.1 documentation".
  54. "NVIDIA Bringing up Open-Source Volta GPU Support for Their Xavier SoC".
  55. "NVIDIA Ada Lovelace Architecture".
  56. Dissecting the Turing GPU Architecture through Microbenchmarking
  57. "H.1. Features and Technical Specifications – Table 13. Feature Support per Compute Capability". docs.nvidia.com. Retrieved 2020-09-23.
  58. "CUDA C++ Programming Guide".
  59. Fused-Multiply-Add, actually executed, Dense Matrix
  60. as SASS since 7.5, as PTX since 8.0
  61. ^ unofficial support in SASS
  62. "Technical brief. NVIDIA Jetson AGX Orin Series" (PDF). nvidia.com. Retrieved 5 September 2023.
  63. "NVIDIA Ampere GA102 GPU Architecture" (PDF). nvidia.com. Retrieved 5 September 2023.
  64. Luo, Weile; Fan, Ruibo; Li, Zeyu; Du, Dayou; Wang, Qiang; Chu, Xiaowen (2024). "Benchmarking and Dissecting the Nvidia Hopper GPU Architecture". arXiv:2402.13499v1 .
  65. "Datasheet NVIDIA A40" (PDF). nvidia.com. Retrieved 27 April 2024.
  66. "NVIDIA AMPERE GA102 GPU ARCHITECTURE" (PDF). 27 April 2024.
  67. "Datasheet NVIDIA L40" (PDF). 27 April 2024.
  68. In the Whitepapers the Tensor Core cube diagrams represent the Dot Product Unit Width into the height (4 FP16 for Volta and Turing, 8 FP16 for A100, 4 FP16 for GA102, 16 FP16 for GH100). The other two dimensions represent the number of Dot Product Units (4x4 = 16 for Volta and Turing, 8x4 = 32 for Ampere and Hopper). The resulting gray blocks are the FP16 FMA operations per cycle. Pascal without Tensor core is only shown for speed comparison as is Volta V100 with non-FP16 datatypes.
  69. "NVIDIA Turing Architecture Whitepaper" (PDF). nvidia.com. Retrieved 5 September 2023.
  70. "NVIDIA Tensor Core GPU" (PDF). nvidia.com. Retrieved 5 September 2023.
  71. "NVIDIA Hopper Architecture In-Depth". 22 March 2022.
  72. ^ shape x converted operand size, e.g. 2 tensor cores x 4x4x4xFP16/cycle = 256 Bytes/cycle
  73. ^ = product first 3 table rows
  74. ^ = product of previous 2 table rows; shape: e.g. 8x8x4xFP16 = 512 Bytes
  75. Sun, Wei; Li, Ang; Geng, Tong; Stuijk, Sander; Corporaal, Henk (2023). "Dissecting Tensor Cores via Microbenchmarks: Latency, Throughput and Numeric Behaviors". IEEE Transactions on Parallel and Distributed Systems. 34 (1): 246–261. arXiv:2206.02874. doi:10.1109/tpds.2022.3217824. S2CID 249431357.
  76. "Parallel Thread Execution ISA Version 7.7".
  77. Raihan, Md Aamir; Goli, Negar; Aamodt, Tor (2018). "Modeling Deep Learning Accelerator Enabled GPUs". arXiv:1811.08309 .
  78. "NVIDIA Ada Lovelace Architecture".
  79. ^ Jia, Zhe; Maggioni, Marco; Smith, Jeffrey; Daniele Paolo Scarpazza (2019). "Dissecting the NVidia Turing T4 GPU via Microbenchmarking". arXiv:1903.07486 .
  80. Burgess, John (2019). "RTX ON – The NVIDIA TURING GPU". 2019 IEEE Hot Chips 31 Symposium (HCS). pp. 1–27. doi:10.1109/HOTCHIPS.2019.8875651. ISBN 978-1-7281-2089-8. S2CID 204822166.
  81. Burgess, John (2019). "RTX ON – The NVIDIA TURING GPU". 2019 IEEE Hot Chips 31 Symposium (HCS). pp. 1–27. doi:10.1109/HOTCHIPS.2019.8875651. ISBN 978-1-7281-2089-8. S2CID 204822166.
  82. dependent on device
  83. ^ "Tegra X1". 9 January 2015.
  84. NVIDIA H100 Tensor Core GPU Architecture
  85. H.1. Features and Technical Specifications – Table 14. Technical Specifications per Compute Capability
  86. NVIDIA Hopper Architecture In-Depth
  87. can only execute 160 integer instructions according to programming guide
  88. 128 according to . 64 from FP32 + 64 separate units?
  89. 64 by FP32 cores and 64 by flexible FP32/INT cores.
  90. "CUDA C++ Programming Guide".
  91. 32 FP32 lanes combine to 16 FP64 lanes. Maybe lower depending on model.
  92. only supported by 16 FP32 lanes, they combine to 4 FP64 lanes
  93. ^ depending on model
  94. Effective speed, probably over FP32 ports. No description of actual FP64 cores.
  95. Can also be used for integer additions and comparisons
  96. 2 clock cycles/instruction for each SM partition Burgess, John (2019). "RTX ON – The NVIDIA TURING GPU". 2019 IEEE Hot Chips 31 Symposium (HCS). pp. 1–27. doi:10.1109/HOTCHIPS.2019.8875651. ISBN 978-1-7281-2089-8. S2CID 204822166.
  97. Durant, Luke; Giroux, Olivier; Harris, Mark; Stam, Nick (May 10, 2017). "Inside Volta: The World's Most Advanced Data Center GPU". Nvidia developer blog.
  98. The schedulers and dispatchers have dedicated execution units unlike with Fermi and Kepler.
  99. Dispatching can overlap concurrently, if it takes more than one cycle (when there are less execution units than 32/SM Partition)
  100. Can dual issue MAD pipe and SFU pipe
  101. No more than one scheduler can issue 2 instructions at once. The first scheduler is in charge of warps with odd IDs. The second scheduler is in charge of warps with even IDs.
  102. ^ shared memory only, no data cache
  103. ^ shared memory separate, but L1 includes texture cache
  104. "H.6.1. Architecture". docs.nvidia.com. Retrieved 2019-05-13.
  105. "Demystifying GPU Microarchitecture through Microbenchmarking" (PDF).
  106. ^ Jia, Zhe; Maggioni, Marco; Staiger, Benjamin; Scarpazza, Daniele P. (2018). "Dissecting the NVIDIA Volta GPU Architecture via Microbenchmarking". arXiv:1804.06826 .
  107. Jia, Zhe; Maggioni, Marco; Smith, Jeffrey; Daniele Paolo Scarpazza (2019). "Dissecting the NVidia Turing T4 GPU via Microbenchmarking". arXiv:1903.07486 .
  108. "Dissecting the Ampere GPU Architecture through Microbenchmarking".
  109. Note that Jia, Zhe; Maggioni, Marco; Smith, Jeffrey; Daniele Paolo Scarpazza (2019). "Dissecting the NVidia Turing T4 GPU via Microbenchmarking". arXiv:1903.07486 . disagrees and states 2 KiB L0 instruction cache per SM partition and 16 KiB L1 instruction cache per SM
  110. "asfermi Opcode". GitHub.
  111. ^ for access with texture engine only
  112. 25% disabled on RTX 4090
  113. "I.7. Compute Capability 8.x". docs.nvidia.com. Retrieved 2022-10-12.
  114. "Appendix F. Features and Technical Specifications" (PDF). (3.2 MiB), page 148 of 175 (Version 5.0 October 2012).
  115. "nVidia CUDA Bioinformatics: BarraCUDA". BioCentric. 2019-07-19. Retrieved 2019-10-15.
  116. "Part V: Physics Simulation". NVIDIA Developer. Retrieved 2020-09-11.
  117. "oneAPI Programming Model". oneAPI.io. Retrieved 2024-07-27.
  118. "Specifications | oneAPI". oneAPI.io. Retrieved 2024-07-27.
  119. "oneAPI Specification — oneAPI Specification 1.3-rev-1 documentation". oneapi-spec.uxlfoundation.org. Retrieved 2024-07-27.
  120. "Exclusive: Behind the plot to break Nvidia's grip on AI by targeting software". Reuters. Retrieved 2024-04-05.
  121. "Question: What does ROCm stand for? · Issue #1628 · RadeonOpenCompute/ROCm". Github.com. Retrieved January 18, 2022.

Further reading

External links

Nvidia
GeForce (List of GPUs)
Fixed pixel pipeline
Pre-GeForce
Vertex and pixel shaders
  • GeForce 3
  • 4 Ti
  • FX
  • 6
  • 7
    • Unified shaders
    Unified shaders & NUMA
    Ray tracing & Tensor Cores
    Software and technologies
    Multimedia acceleration
    Software
    Technologies
    GPU microarchitectures
    Other products
    Graphics Workstation cards
    GPGPU
    Console components
    Nvidia Shield
    SoCs and embedded
    CPUs
    Computer chipsets
    Company
    Key people
    Acquisitions
    Processor technologies
    Models
    Architecture
    Instruction set
    architectures
    Types
    Instruction
    sets
    Execution
    Instruction pipelining
    Hazards
    Out-of-order
    Speculative
    Parallelism
    Level
    Multithreading
    Flynn's taxonomy
    Processor
    performance
    Types
    By application
    Systems
    on chip
    Hardware
    accelerators
    Word size
    Core count
    Components
    Functional
    units
    Logic
    Registers
    Control unit
    Datapath
    Circuitry
    Power
    management
    Related
    Parallel computing
    General
    Levels
    Multithreading
    Theory
    Elements
    Coordination
    Programming
    Hardware
    APIs
    Problems
    Categories:
    CUDA: Difference between revisions Add topic