Nvidia Tesla

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 is the former name for a line of products developed by Nvidia 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.

Nvidia Tesla

Manufacturer	Nvidia
Introduced	May 2, 2007; 18 years ago (2007-05-02)
Discontinued	The Tesla branding was discontinued in May 2020; 5 years ago (2020-05) - now branded as Nvidia Data Center GPUs
Type	General purpose graphics cards

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 Nvidia Data Center GPUs^[2] as in the Ampere-based A100 GPU.^[3]

Nvidia DGX servers feature Nvidia GPGPUs.

Overview

Nvidia Tesla C2075

Offering computational power much greater than traditional microprocessors, the Tesla products targeted the high-performance computing market.^[4] As of 2012^[update], 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.^{[citation needed]} 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.^[5]

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.^[6]

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.^[7]

Specifications

Model	Microarchitecture	Launch	Core	Core clock (MHz)	Shaders			Memory					Processing power (TFLOPS)[a]			CUDA compute capability[b]	TDP (W)	Notes, form factor
					CUDA cores (total)	Base clock (MHz)	Max boost clock (MHz)[c]	Bus type	Bus width (bit)	Size (GB)	Clock (MT/s)	Bandwidth (GB/s)	Half precision Tensor Core FP32 Accumulate	Single precision (MAD or FMA)	Double precision (FMA)
C870 GPU Computing Module[d]	Tesla	May 2, 2007	1× G80	600	128	1350	—	GDDR3	384	1.5	1600	76.8	No	0.3456	No	1.0	170.9	Internal PCIe GPU (full-height, dual-slot)
D870 Deskside Computer[d]		May 2, 2007	2× G80	600	256	1350	—	GDDR3	2× 384	2× 1.5	1600	2× 76.8	No	0.6912	No	1.0	520	Deskside or 3U rack-mount external GPUs
S870 GPU Computing Server[d]		May 2, 2007	4× G80	600	512	1350	—	GDDR3	4× 384	4× 1.5	1600	4× 76.8	No	1.382	No	1.0		1U rack-mount external GPUs, connect via 2× PCIe (×16)
C1060 GPU Computing Module[e]		April 9, 2009	1× GT200	602	240	1296[9]	—	GDDR3	512	4	1600	102.4	No	0.6221	0.07776	1.3	187.8	Internal PCIe GPU (full-height, dual-slot)
S1070 GPU Computing Server "400 configuration"[e]		June 1, 2008	4× GT200	602	960	1296	—	GDDR3	4× 512	4× 4	1538.4	4× 98.5	No	2.488	0.311	1.3	800	1U rack-mount external GPUs, connect via 2× PCIe (×8 or ×16)
S1070 GPU Computing Server "500 configuration"[e]		June 1, 2008				1440	—						No	2.765	0.3456
S1075 GPU Computing Server[e][10]		June 1, 2008	4× GT200	602	960	1440	—	GDDR3	4× 512	4× 4	1538	4× 98.5	No	2.765	0.3456	1.3		1U rack-mount external GPUs, connect via 1× PCIe (×8 or ×16)
Quadro Plex 2200 D2 Visual Computing System[f]		July 25, 2008	2× GT200GL	648	480	1296	—	GDDR3	2× 512	2× 4	1600	2× 102.4	No	1.244	0.1555	1.3		Deskside or 3U rack-mount external GPUs with 4 dual-link DVI outputs
Quadro Plex 2200 S4 Visual Computing System[f]		July 25, 2008	4× GT200GL	648	960	1296	—	GDDR3	4× 512	4× 4	1600	4× 102.4	No	2.488	0.311	1.3	1200	1U rack-mount external GPUs, connect via 2× PCIe (×8 or ×16)
C2050 GPU Computing Module[11]	Fermi	July 25, 2011	1× GF100	575	448	1150	—	GDDR5	384	3[g]	3000	144	No	1.030	0.5152	2.0	247	Internal PCIe GPU (full-height, dual-slot)
M2050 GPU Computing Module[12]		July 25, 2011					—				3092	148.4	No				225
C2070 GPU Computing Module[11]		July 25, 2011	1× GF100	575	448	1150	—	GDDR5	384	6[g]	3000	144	No	1.030	0.5152	2.0	247	Internal PCIe GPU (full-height, dual-slot)
C2075 GPU Computing Module[13]		July 25, 2011					—				3000	144	No				225
M2070/M2070Q GPU Computing Module[14]		July 25, 2011					—				3132	150.3	No				225
M2090 GPU Computing Module[15]		July 25, 2011	1× GF110	650	512	1300	—	GDDR5	384	6[g]	3700	177.6	No	1.331	0.6656	2.0	225	Internal PCIe GPU (full-height, dual-slot)
S2050 GPU Computing Server		July 25, 2011	4× GF100	575	1792	1150	—	GDDR5	4× 384	4× 3[g]	3000	4× 148.4	No	4.122	2.061	2.0	900	1U rack-mount external GPUs, connect via 2× PCIe (×8 or ×16)
S2070 GPU Computing Server		July 25, 2011					—			4× 6[g]			No
K10 GPU accelerator[16]	Kepler	May 1, 2012	2× GK104	—	3072	745	?	GDDR5	2× 256	2× 4	5000	2× 160	No	4.577	0.1907	3.0	225	Internal PCIe GPU (full-height, dual-slot)
K20 GPU accelerator[17][18]		November 12, 2012	1× GK110	—	2496	706	758	GDDR5	320	5	5200	208	No	3.524	1.175	3.5	225	Internal PCIe GPU (full-height, dual-slot)
K20X GPU accelerator[19]		November 12, 2012	1× GK110	—	2688	732	?	GDDR5	384	6	5200	250	No	3.935	1.312	3.5	235	Internal PCIe GPU (full-height, dual-slot)
K40 GPU accelerator[20]		October 8, 2013	1× GK110B	—	2880	745	875	GDDR5	384	12[g]	6000	288	No	4.291–5.040	1.430–1.680	3.5	235	Internal PCIe GPU (full-height, dual-slot)
K80 GPU accelerator[21]		November 17, 2014	2× GK210	—	4992	560	875	GDDR5	2× 384	2× 12	5000	2× 240	No	5.591–8.736	1.864–2.912	3.7	300	Internal PCIe GPU (full-height, dual-slot)
M4 GPU accelerator[22][23]	Maxwell	November 10, 2015	1× GM206	—	1024	872	1072	GDDR5	128	4	5500	88	No	1.786–2.195	0.05581–0.06861	5.2	50–75	Internal PCIe GPU (half-height, single-slot)
M6 GPU accelerator[24]		August 30, 2015	1× GM204-995-A1	—	1536	722	1051	GDDR5	256	8	4600	147.2	No	2.218–3.229	0.0693–0.1009	5.2	75–100	Internal MXM GPU
M10 GPU accelerator[25]		May 18, 2016	4× GM107	—	2560	1033	?	GDDR5	4× 128	4× 8	5188	4× 83	No	5.289	0.1653	5.2	225	Internal PCIe GPU (full-height, dual-slot)
M40 GPU accelerator[23][26]		November 10, 2015	1× GM200	—	3072	948	1114	GDDR5	384	12 or 24	6000	288	No	5.825–6.844	0.182–0.2139	5.2	250	Internal PCIe GPU (full-height, dual-slot)
M60 GPU accelerator[27]		August 30, 2015	2× GM204-895-A1	—	4096	899	1178	GDDR5	2× 256	2× 8	5000	2× 160	No	7.365–9.650	0.2301–0.3016	5.2	225–300	Internal PCIe GPU (full-height, dual-slot)
P4 GPU accelerator[28]	Pascal	September 13, 2016	1× GP104	—	2560	810	1063	GDDR5	256	8	6000	192.0	No	4.147–5.443	0.1296–0.1701	6.1	50–75	PCIe card
P6 GPU accelerator[29][30]		March 24, 2017	1× GP104-995-A1	—	2048	1012	1506	GDDR5	256	16	3003	192.2	No	6.169	0.1928	6.1	90	MXM card
P40 GPU accelerator[28]		September 13, 2016	1× GP102	—	3840	1303	1531	GDDR5	384	24	7200	345.6	No	10.01–11.76	0.3127–0.3674	6.1	250	PCIe card
P100 GPU accelerator (mezzanine)[31][32]		April 5, 2016	1× GP100-890-A1	—	3584	1328	1480	HBM2	4096	16	1430	732	No	9.519–10.61	4.760–5.304	6.0	300	SXM card
P100 GPU accelerator (16 GB card)[33]		June 20, 2016	1× GP100	—		1126	1303						No	8.071‒9.340	4.036‒4.670		250	PCIe card
P100 GPU accelerator (12 GB card)[33]		June 20, 2016		—					3072	12		549	No	8.071‒9.340	4.036‒4.670
V100 GPU accelerator (mezzanine)[34][35][36]	Volta	May 10, 2017	1× GV100-895-A1	—	5120	Un­known	1455	HBM2	4096	16 or 32	1750	900	119.2	14.90	7.450	7.0	300	SXM card
V100 GPU accelerator (PCIe card)[34][35][36]		June 21, 2017	1× GV100	—		Un­known	1370						112.2	14.03	7.014		250	PCIe card
V100 GPU accelerator (PCIe FHHL card)		March 27, 2018	1× GV100	—		937	1290			16	1620	829.4	105.7	13.21	6.605		250	PCIe FHHL card
T4 GPU accelerator (PCIe card)[37][38]	Turing	September 12, 2018	1× TU104-895-A1	—	2560	585	1590	GDDR6	256	16	5000	320	64.8	8.1	Un­known	7.5	70	PCIe card
A2 GPU accelerator (PCIe card)[39]	Ampere	November 10, 2021	1× GA107	—	1280	1440	1770	GDDR6	128	16	6252	200	18.12	4.531	0.14	8.6	40–60	PCIe card (half height, single-slot)
A10 GPU accelerator (PCIe card)[40]		April 12, 2021	1× GA102-890-A1	—	9216	885	1695	GDDR6	384	24	6252	600	125.0	31.24	0.976	8.6	150	PCIe card (single-slot)
A16 GPU accelerator (PCIe card)[41]		April 12, 2021	4× GA107	—	4× 1280	885	1695	GDDR6	4× 128	4× 16	7242	4× 200	4x 18.43	4× 4.608	1.085	8.6	250	PCIe card (dual-slot)
A30 GPU accelerator (PCIe card)[42]		April 12, 2021	1× GA100	—	3584	930	1440	HBM2	3072	24	1215	933.1	165.1	10.32	5.161	8.0	165	PCIe card (dual-slot)
A40 GPU accelerator (PCIe card)[43]		October 5, 2020	1× GA102	—	10,752	1305	1740	GDDR6	384	48	7248	695.8	149.7	37.42	1.168	8.6	300	PCIe card (dual-slot)
A100 GPU accelerator (PCIe card)[44][45]		May 14, 2020[46]	1× GA100-883AA-A1	—	6912	765	1410	HBM2	5120	40 or 80	1215	1555	312.0	19.5	9.7	8.0	250	PCIe card (dual-slot)
H100 GPU accelerator (PCIe card)[47]	Hopper	March 22, 2022[48]	1× GH100[49]	—	14,592	1065	1755 CUDA 1620 TC	HBM2E	5120	80	1000	2039	756.4	51.2	25.6	9.0	350	PCIe card (dual-slot)
H100 GPU accelerator (SXM card)				—	16,896	1065	1980 CUDA 1830 TC	HBM3	5120	64 or 80 or 96	1500	3352	989.4	66.9	33.5	9.0	700	SXM card
H200 GPU accelerator (PCIe card)[50]		November 18, 2024[51]	1× GH100	—	16,896	1365	1785	HBM3E	5120	141	1313	3360	—	60.32	30.16	9.0	600	PCIe card (dual-slot)
H200 GPU accelerator (SXM card)				—	16,896	1590	1980	HBM3E	5120	141	1313	3360	—	66.91	33.45	9.0	700	SXM card
H800 GPU accelerator (SXM card)		March 21, 2023[52]	1× GH100	—	16,896	1095	1755	HBM3	5120	80	1313	3360	—	59.3	29.65	9.0	700	SXM card
L40 GPU accelerator[53]	Ada Lovelace	October 13, 2022	1× AD102[54]	—	18,176	735	2490	GDDR6	384	48	2250	864	362.1	90.52	1.414	8.9	300	PCIe card (dual-slot)
L4 GPU accelerator[55][56]		March 21, 2023[57]	1x AD104[58]	—	7424	795	2040	GDDR6	192	24	1563	300	121.0	30.3	0.49	8.9	72	HHHL single slot PCIe card

• v
• t
• e

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.^[8]
4. Specifications not specified by Nvidia assumed to be based on the GeForce 8800 GTX
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.)

See also

• List of Nvidia graphics processing units
• Nvidia Tesla Personal Supercomputer
• Ampere (microarchitecture)
• Fastra II