GPU Specs

No longer do you need to go google a bunch of nvidia datasheets to figure compare performance.

If any of the numbers look sus, just submit a PR.

For cost data on specific major providers, look in costs/, where there's a summary comparing providers, and then you can check out individual ones in their respective markdown files.

---

Specs Table

GPU	Gen	VRAM	Bandwidth	FP64	FP32/TF32	FP16/BF16	FP8	INT8	FP4
B200	Blackwell	192GB	8.0 TB/s	30	1,125	2,250	4,500	4,500	9,000
RTX PRO 6000 Blackwell	Blackwell	96GB	1.8 TB/s	1.97	126	126	252	504	1,008
RTX 5090	Blackwell	32GB	1.8 TB/s	1.64	104.8	209.5	419	419	838
H200 SXM	Hopper	141GB	4.8 TB/s	67	989	989	1,979	3,958	-
H100 SXM	Hopper	80GB	3.35 TB/s	67	989	989	1,979	3,958	-
H100 PCIe	Hopper	80GB	2.0 TB/s	51	756	756	1,513	3,026	-
H100 NVL	Hopper	94GB	3.9 TB/s	67	989	989	1,979	3,958	-
RTX 6000 Ada	Ada Lovelace	48GB	960 GB/s	-	91.1	91.1	182	364	-
RTX 4090	Ada Lovelace	24GB	1.0 TB/s	-	82.6	82.6	165	330	-
RTX 4000 Ada	Ada Lovelace	20GB	320 GB/s	-	26.5	26.5	53	106	-
RTX 2000 Ada	Ada Lovelace	16GB	288 GB/s	-	12	12	24	48	-
L40S	Ada Lovelace	48GB	864 GB/s	-	91.6	91.6	183	366	-
L40	Ada Lovelace	48GB	864 GB/s	-	90.5	90.5	181	362	-
L4	Ada Lovelace	24GB	300 GB/s	-	30.3	30.3	60.6	121	-
A100 80GB	Ampere	80GB	2.0 TB/s	9.7	156	312	-	624	-
A100 40GB	Ampere	40GB	1.6 TB/s	9.7	156	312	-	624	-
A40	Ampere	48GB	696 GB/s	-	37.4	37.4	-	74.8	-
RTX A6000	Ampere	48GB	768 GB/s	-	38.7	38.7	-	77.4	-
RTX A5000	Ampere	24GB	768 GB/s	-	27.7	27.7	-	55.4	-
RTX A4500	Ampere	20GB	640 GB/s	-	23.7	23.7	-	47.4	-
RTX A4000	Ampere	16GB	448 GB/s	-	19.2	19.2	-	38.4	-
RTX 3090	Ampere	24GB	936 GB/s	-	35.6	35.6	-	71.2	-
RTX 3090 Ti	Ampere	24GB	1.0 TB/s	-	40	40	-	80	-
RTX 3080	Ampere	10GB	760 GB/s	-	29.8	29.8	-	59.6	-
V100	Volta	16GB	900 GB/s	7.8	15.7	31.4	-	-	-
T4	Turing	16GB	320 GB/s	-	8.1	65	-	130	-
A10G	Ampere	24GB	600 GB/s	-	31.2	31.2	-	62.4	-

Notes

• All TFLOPS values are theoretical maximum performance (non-sparse unless specified)
• Dash (-) means not supported by that architecture
• FP16/BF16 values shown are for Tensor Core performance where available
• RTX 5090 specifications based on Blackwell architecture with 5th gen Tensor Cores
• H100 SXM uses HBM3 memory with 3.35 TB/s bandwidth (not 3.0 TB/s)
• RTX PRO 6000 Blackwell features 96GB GDDR7 and targets professional workstations
• Consumer Blackwell GPUs (RTX 50 series) support FP4 precision
• Professional Ada GPUs (RTX 6000 Ada, L40S, etc.) use standard FP16 without sparsity
• Ampere A100 introduced support for structured sparsity (2:4) in Tensor Core operations