Nvidia Tesla

Nvidia Tesla was the name of Nvidia's line of products targeted at stream processing or general-purpose graphics processing units (GPGPU), named after pioneering electrical engineer Nikola Tesla. Its products began using GPUs from the G80 series, and have continued to accompany the release of new chips. They are programmable using the CUDA or OpenCL APIs.

This article is about GPGPU cards. For the GPU microarchitecture, see Tesla (microarchitecture).
"Tesla P100" redirects here. For the line of performance cars by Tesla Motors (P100D), see Tesla Model S and Tesla Model X.

Nvidia Tesla

The Nvidia Tesla product line competed with AMD's Radeon Instinct and Intel Xeon Phi lines of deep learning and GPU cards.

Nvidia retired the Tesla brand in May 2020, reportedly because of potential confusion with the brand of cars.[1] Its new GPUs are branded Ampere, as in the Ampere A100 GPU.[2]

Overview
Nvidia Tesla C2075

Offering computational power much greater than traditional microprocessors, the Tesla products targeted the high-performance computing market.[3] As of 2012, Nvidia Teslas power some of the world's fastest supercomputers, including Summit at Oak Ridge National Laboratory and Tianhe-1A, in Tianjin, China.

Tesla cards have four times the double precision performance of a Fermi-based Nvidia GeForce card of similar single precision performance. Unlike Nvidia's consumer GeForce cards and professional Nvidia Quadro cards, Tesla cards were originally unable to output images to a display. However, the last Tesla C-class products included one Dual-Link DVI port.[4]

As part of Project Denver, Nvidia intends to embed ARMv8 processor cores in its GPUs.[5] This will be a 64-bit follow-up to the 32-bit Tegra chips.

The Tesla P100 uses TSMC's 16 nanometer FinFET semiconductor manufacturing process, which is more advanced than the 28-nanometer process previously used by AMD and Nvidia GPUs between 2012 and 2016. The P100 also uses Samsung's HBM2 memory.[6]

Applications

Tesla products are primarily used in simulations and in large-scale calculations (especially floating-point calculations), and for high-end image generation for professional and scientific fields.[7]

In 2013, the defense industry accounted for less than one-sixth of Tesla sales, but Sumit Gupta predicted increasing sales to the geospatial intelligence market.[8]

Specifications
Model Micro-
architecture 		Launch Chips Core clock
(MHz)		 Shaders Memory Processing power (GFLOPS)[lower-alpha 1] CUDA
compute
ability[lower-alpha 2]		 TDP
(watts)		 Notes, form_factor
Cuda cores
(total)		 Base clock (MHz) Max boost
clock (MHz)[lower-alpha 3]		 Bus type Bus width
(bit)		 Size
(GB)		 Clock
(MT/s)		 Bandwidth
(GB/s)		 Single precision
(MAD+MUL)		 Single precision
(MAD or FMA)		 Double precision
(FMA)
UnitsMHzMHzW
C870 GPU Computing Module[lower-alpha 4] Tesla May 2, 2007 1× G80 600 128 1350 N/A GDDR3 384 1.5 1600 76.8 518.4 345.6 No 1.0 170.9 Internal PCIe GPU (full-height, dual-slot)
D870 Deskside Computer[lower-alpha 4] May 2, 2007 2× G80 600 256 1350 N/A GDDR3 2× 384 2× 1.5 1600 2× 76.8 1036.8 691.2 No 1.0 520 Deskside or 3U rack-mount external GPUs
S870 GPU Computing Server[lower-alpha 4] May 2, 2007 4× G80 600 512 1350 N/A GDDR3 4× 384 4× 1.5 1600 4× 76.8 2073.6 1382.4 No 1.0 1U rack-mount external GPUs, connect via 2× PCIe (×16)
C1060 GPU Computing Module[lower-alpha 5] April 9, 2009 1× GT200 602 240 1296[10] N/A GDDR3 512 4 1600 102.4 933.12 622.08 77.76 1.3 187.8 Internal PCIe GPU (full-height, dual-slot)
S1070 GPU Computing Server "400 configuration"[lower-alpha 5] June 1, 2008 4× GT200 602 960 1296 N/A GDDR3 4× 512 4× 4 1538.4 4× 98.5 3732.5 2488.3 311.0 1.3 800 1U rack-mount external GPUs, connect via 2× PCIe (×8 or ×16)
S1070 GPU Computing Server "500 configuration"[lower-alpha 5] 1440 N/A 4147.2 2764.8 345.6
S1075 GPU Computing Server[lower-alpha 5][11] June 1, 2008 4× GT200 602 960 1440 N/A GDDR3 4× 512 4× 4 1538.4 4× 98.5 4147.2 2764.8 345.6 1.3 1U rack-mount external GPUs, connect via 1× PCIe (×8 or ×16)
Quadro Plex 2200 D2 Visual Computing System[lower-alpha 6] 2× GT200GL 648 480 1296 N/A GDDR3 2× 512 2× 4 1600 2× 102.4 1866.2 1244.2 155.5 1.3 Deskside or 3U rack-mount external GPUs with 4 dual-link DVI outputs
Quadro Plex 2200 S4 Visual Computing System[lower-alpha 6] 4× GT200GL 648 960 1296 N/A GDDR3 4× 512 4× 4 1600 4× 102.4 3732.5 2488.3 311.0 1.3 1200 1U rack-mount external GPUs, connect via 2× PCIe (×8 or ×16)
C2050 GPU Computing Module[12] Fermi July 25, 2011 1× GF100 575 448 1150 N/A GDDR5 384 3[lower-alpha 7] 3000 144 No 1030.4 515.2 2.0 247 Internal PCIe GPU (full-height, dual-slot)
M2050 GPU Computing Module[13] July 25, 2011 N/A 3092 148.4 No 225
C2070 GPU Computing Module[12] July 25, 2011 1× GF100 575 448 1150 N/A GDDR5 384 6[lower-alpha 7] 3000 144 No 1030.4 515.2 2.0 247 Internal PCIe GPU (full-height, dual-slot)
C2075 GPU Computing Module[14] July 25, 2011 N/A 3000 144 No 225
M2070/M2070Q GPU Computing Module[15] July 25, 2011 N/A 3132 150.336 No 225
M2090 GPU Computing Module[16] July 25, 2011 1× GF110 650 512 1300 N/A GDDR5 384 6[lower-alpha 7] 3700 177.6 No 1331.2 665.6 2.0 225 Internal PCIe GPU (full-height, dual-slot)
S2050 GPU Computing Server July 25, 2011 4× GF100 575 1792 1150 N/A GDDR5 4× 384 4× 3[lower-alpha 7] 3 4× 148.4 No 4121.6 2060.8 2.0 900 1U rack-mount external GPUs, connect via 2× PCIe (×8 or ×16)
S2070 GPU Computing Server N/A 4× 6[lower-alpha 7] No
K10 GPU accelerator[17] Kepler May 1, 2012 2× GK104 N/A 3072 745 ? GDDR5 2× 256 2× 4 5000 2× 160 No 4577 190.7 3.0 225 Internal PCIe GPU (full-height, dual-slot)
K20 GPU accelerator[18][19] November 12, 2012 1× GK110 N/A 2496 706 758 GDDR5 320 5 5200 208 No 3524 1175 3.5 225 Internal PCIe GPU (full-height, dual-slot)
K20X GPU accelerator[20] November 12, 2012 1× GK110 N/A 2688 732 ? GDDR5 384 6 5200 250 No 3935 1312 3.5 235 Internal PCIe GPU (full-height, dual-slot)
K40 GPU accelerator[21] October 8, 2013 1× GK110B N/A 2880 745 875 GDDR5 384 12[lower-alpha 7] 6000 288 No 4291–5040 1430–1680 3.5 235 Internal PCIe GPU (full-height, dual-slot)
K80 GPU accelerator[22] November 17, 2014 2× GK210 N/A 4992 560 875 GDDR5 2× 384 2× 12 5000 2× 240 No 5591–8736 1864–2912 3.7 300 Internal PCIe GPU (full-height, dual-slot)
M4 GPU accelerator[23][24] Maxwell November 10, 2015 1× GM206 N/A 1024 872 1072 GDDR5 128 4 5500 88 No 1786–2195 55.81–68.61 5.2 50–75 Internal PCIe GPU (half-height, single-slot)
M6 GPU accelerator[25] August 30, 2015 1× GM204 N/A 1536 722 1051 GDDR5 256 8 4600 147.2 No 2218–3229 69.3–100.9 5.2 75–100 Internal MXM GPU
M10 GPU accelerator[26] 4× GM107 N/A 2560 1033 ? GDDR5 4× 128 4× 8 5188 4× 83 No 5289 165.3 5.2 225 Internal PCIe GPU (full-height, dual-slot)
M40 GPU accelerator[24][27] November 10, 2015 1× GM200 N/A 3072 948 1114 GDDR5 384 12 6000 288 No 5825–6844 182.0–213.9 5.2 250 Internal PCIe GPU (full-height, dual-slot)
M60 GPU accelerator[28] August 30, 2015 2× GM204 N/A 4096 899 1178 GDDR5 2× 256 2× 8 5000 2× 160 No 7365–9650 230.1–301.6 5.2 225–300 Internal PCIe GPU (full-height, dual-slot)
P4 GPU accelerator[29] Pascal September 13, 2016 1× GP104 N/A 2560 810 1063 GDDR5 256 8 6000 192.0 No 4147–5443 129.6–170.1 6.1 50-75 PCIe card
P6 GPU accelerator[30][31] March 24, 2017 1× GP104-995 N/A 2048 1012 1506 GDDR5 256 16 3003 192.2 No 6169 192.8 6.1 90 MXM card
P40 GPU accelerator[29] September 13, 2016 1× GP102 N/A 3840 1303 1531 GDDR5 384 24 7200 345.6 No 10007–11758 312.7–367.4 6.1 250 PCIe card
P100 GPU accelerator (mezzanine)[32][33] April 5, 2016 1× GP100 N/A 3584 1328 1480 HBM2 4096 16 1430 732 No 9519–10609 4760–5304 6.0 300 NVLink card
P100 GPU accelerator (16 GB card)[34] June 20, 2016 N/A 1126 1303 No 8071‒9340 4036‒4670 250 PCIe card
P100 GPU accelerator (12 GB card)[34] June 20, 2016 N/A 3072 12 549 No 8071‒9340 4036‒4670
V100 GPU accelerator (mezzanine)[35][36][37] Volta 1× GV100 N/A 5120 Unknown 1455 HBM2 4096 16 or 32 1750 900 No 14899 7450 7.0 300 NVlink card
V100 GPU accelerator (PCIe card)[35][36][37] June 21, 2017 N/A Unknown 1370 No 14028 7014 250 PCIe card
T4 GPU accelerator (PCIe card)[38][39] Turing September 12, 2018 1× TU104 N/A 2560 585 1590 GDDR6 256 16 Unknown 320 No 8100 Unknown 7.5 70 PCIe card
Model Micro-
architecture 		Launch Chips Core clock
(MHz)		 Shaders Memory Processing power (GFLOPS)[lower-alpha 1] CUDA
compute
ability		 TDP
(watts)		 Notes, form factor
Cuda cores
(total)		 Base clock (MHz) Max boost
clock (MHz)[lower-alpha 3]		 Bus type Bus width
(bit)		 Size
(GB)		 Clock
(MT/s)		 Bandwidth
(total)
(GB/s)		 Single precision
(MAD+MUL)		 Single precision
(MAD or FMA)		 Double precision
(FMA)

Notes

  1. To calculate the processing power see Tesla (microarchitecture)#Performance, Fermi (microarchitecture)#Performance, Kepler (microarchitecture)#Performance, Maxwell (microarchitecture)#Performance, or Pascal (microarchitecture)#Performance. A number range specifies the minimum and maximum processing power at, respectively, the base clock and maximum boost clock.
  2. Core architecture version according to the CUDA programming guide.
  3. GPU Boost is a default feature that increases the core clock rate while remaining under the card's predetermined power budget. Multiple boost clocks are available, but this table lists the highest clock supported by each card.[9]
  4. Specifications not specified by Nvidia assumed to be based on the GeForce 8800GTX
  5. Specifications not specified by Nvidia assumed to be based on the GeForce GTX 280
  6. Specifications not specified by Nvidia assumed to be based on the Quadro FX 5800
  7. With ECC on, a portion of the dedicated memory is used for ECC bits, so the available user memory is reduced by 12.5%. (e.g. 4 GB total memory yields 3.5 GB of user available memory.)
gollark: The vectorised implementation of firecubez is planned for 2024.
gollark: Vier-zee-ion.
gollark: This is also approximately why I'm against more globalized governance integration: having multiple somewhat independent nations/states/whatever means you can test out different plans in parallel without having to *explicitly* A/B test people, which they dislike.
gollark: I'm not sure if your premise applies in realistic cases.
gollark: Sure.

See also

References
  1. Casas, Alex (19 May 2020). "NVIDIA Drops Tesla Brand To Avoid Confusion With Tesla". Wccftech. Retrieved 8 July 2020.
  2. "NVIDIA A100 GPUs Power the Modern Data Center". NVIDIA. Retrieved 8 July 2020.
  3. "High Performance Computing - Supercomputing with Tesla GPUs".
  4. "Professional Workstation Solutions".
  5. "Nvidia to Integrate ARM Processors in Tesla". 1 November 2012.
  6. Walton, Mark (6 April 2016). "Nvidia unveils first Pascal graphics card, the monstrous Tesla P100". Ars Technica. Retrieved 19 June 2019.
  7. Tesla Technical Brief (PDF)
  8. "Nvidia chases defense, intelligence ISVs with GPUs". www.theregister.com. Retrieved 8 July 2020.
  9. "Nvidia GPU Boost For Tesla" (PDF). January 2014. Retrieved 7 December 2015.
  10. "Tesla C1060 Computing Processor Board" (PDF). Nvidia.com. Retrieved 11 December 2015.
  11. "Difference between Tesla S1070 and S1075". 31 October 2008. Retrieved 29 January 2017. S1075 has one interface card
  12. "Tesla C2050 and Tesla C2070 Computing Processor" (PDF). Nvidia.com. Retrieved 11 December 2015.
  13. "Tesla M2050 and Tesla M2070/M2070Q Dual-Slot Computing Processor Modules" (PDF). Nvidia.com. Retrieved 11 December 2015.
  14. "Tesla C2075 Computing Processor Board" (PDF). Nvidia.com. Retrieved 11 December 2015.
  15. Hand, Randall (23 August 2010). "NVidia Tesla M2050 & M2070/M2070Q Specs OnlineVizWorld.com". VizWorld.com. Retrieved 11 December 2015.
  16. "Tesla M2090 Dual-Slot Computing Processor Module" (PDF). Nvidia.com. Retrieved 11 December 2015.
  17. "Tesla K10 GPU accelerator" (PDF). Nvidia.com. Retrieved 11 December 2015.
  18. "Tesla K20 GPU active accelerator" (PDF). Nvidia.com. Retrieved 11 December 2015.
  19. "Tesla K20 GPU accelerator" (PDF). Nvidia.com. Retrieved 11 December 2015.
  20. "Tesla K20X GPU accelerator" (PDF). Nvidia.com. Retrieved 11 December 2015.
  21. "Tesla K40 GPU accelerator" (PDF). Nvidia.com. Retrieved 11 December 2015.
  22. "Tesla K80 GPU accelerator" (PDF). Images.nvidia.com. Retrieved 11 December 2015.
  23. "Nvidia Announces Tesla M40 & M4 Server Cards - Data Center Machine Learning". Anandtech.com. Retrieved 11 December 2015.
  24. "Accelerating Hyperscale Datacenter Applications with Tesla GPUs | Parallel Forall". Devblogs.nvidia.com. 10 November 2015. Retrieved 11 December 2015.
  25. "Tesla M6" (PDF). Images.nvidia.com. Retrieved 28 May 2016.
  26. "Tesla M10" (PDF). Images.nvidia.com. Retrieved 29 October 2016.
  27. "Tesla M40" (PDF). Images.nvidia.com. Retrieved 11 December 2015.
  28. "Tesla M60" (PDF). Images.nvidia.com. Retrieved 27 May 2016.
  29. Smith, Ryan (13 September 2016). "Nvidia Announces Tesla P40 & Tesla P4 - Network Inference, Big & Small". Anandtech. Retrieved 13 September 2016.
  30. "Tesla P6" (PDF). www.nvidia.com. Retrieved 7 March 2019.
  31. "Tesla P6 Specs". www.techpowerup.com. Retrieved 7 March 2019.
  32. Smith, Ryan (5 April 2016). "Nvidia Announces Tesla P100 Accelerator - Pascal GP100 for HPC". Anandtech.com. Anandtech.com. Retrieved 5 April 2016.
  33. Harris, Mark. "Inside Pascal: Nvidia's Newest Computing Platform". Retrieved 13 September 2016.
  34. Smith, Ryan (20 June 2016). "NVidia Announces PCI Express Tesla P100". Anandtech.com. Retrieved 21 June 2016.
  35. Smith, Ryan (10 May 2017). "The Nvidia GPU Technology Conference 2017 Keynote Live Blog". Anandtech. Retrieved 10 May 2017.
  36. Smith, Ryan (10 May 2017). "NVIDIA Volta Unveiled: GV100 GPU and Tesla V100 Accelerator Announced". Anandtech. Retrieved 10 May 2017.
  37. Oh, Nate (20 June 2017). "NVIDIA Formally Announces V100: Available later this Year". Anandtech.com. Retrieved 20 June 2017.
  38. "NVIDIA TESLA T4 TENSOR CORE GPU". NVIDIA. Retrieved 17 October 2018.
  39. "NVIDIA Tesla T4 Tensor Core Product Brief" (PDF). www.nvidia.com. Retrieved 10 July 2019.
This article is issued from Wikipedia. The text is licensed under Creative Commons - Attribution - Sharealike. Additional terms may apply for the media files.