常见GPU算力比较（X100计算卡，GX100）

比较对象：

        V100 SXM2、V100 PCIe、V100S PCIe

        A100 40GB PCIe、A100 80GB PCIe、A100 40GB SXM、A100 80GB SXM

        H100 SXM5、H100 PCIe

陪跑：4090

一、硬件参数

	V100 SXM2	V100 PCIe	V100S PCIe	A100 40GB PCIe	A100 80GB PCIe	A100 40GB SXM	A100 80GB SXM	H100 SXM5	H100 PCIe	4090
核心	GV100	GV100	GV100	GA100	GA100	GA100	GA100	GH100	GH100	AD102-300
架构	Volta	Volta	Volta	Ampere	Ampere	Ampere	Ampere	Hopper	Hopper	Ada Lovelace
SM	80	80	80	108	108	108	108	132	114	128
CUDA Cores / SM	64	64	64	64	64	64	64	128	128	128
CUDA Cores / GPU	5120	5120	5120	6912	6912	6912	6912	16896	14592	16384
FP32 Cores / SM	64	64	64	64	64	64	64	128	128	128*
FP32 Cores / GPU	5120	5120	5120	6912	6912	6912	6912	16896	14592	16384
FP64 Cores / SM	32	32	32	32	32	32	32	64	64	2
FP64 Cores / GPU	2560	2560	2560	3456	3456	3456	3456	8448	7296	256
INT32 Cores / SM	64	64	64	64	64	64	64	64	64	64*
INT32 Cores / GPU	5120	5120	5120	6912	6912	6912	6912	8448	7296	8192
Tensor Core	1st	1st	1st	3rd	3rd	3rd	3rd	4th	4th	4th
Tensor Cores / SM	8	8	8	4	4	4	4	4	4	4
Tensor Cores / GPU	640	640	640	432	432	432	432	528	456	512
GPU 加速频率 (MHz)	1530	1380	1597	1410	1410	1410	1410	1830 / 1980**	1620 / 1755**	2520
显存	16 / 32 GB HBM2	16 / 32 GB HBM2	32 GB HBM2	40 GB HBM2	80 GB HBM2e	40 GB HBM2	80 GB HBM2e	80 GB HBM3	80 GB HBM2e	24 GB GDDR6X
显存位宽 (bit)	4096	4096	4096	5120	5120	5120	5120	5120	5120	384
显存带宽 (GBps)	897	897	1133	1555	1935	1555	2039	3352	2039	1008
一缓 (KB per SM)	128	128	128	192	192	192	192	256	256	128
二缓 (MB)	6	6	6	40	40	40	40	50	50	72
接口	SXM2	PCIe 3.0x16	PCIe 3.0x16	PCIe 4.0x16	PCIe 4.0x16	SXM4	SXM4	SXM5	PCIe 5.0x16	PCIe 4.0x16
TDP (W)	300	250	250	250	300	400	400	700	350	450
制程	TSMC 12nm FFN	TSMC 12nm FFN	TSMC 12nm FFN	TSMC N7 (7nm)	TSMC N7 (7nm)	TSMC N7 (7nm)	TSMC N7 (7nm)	TSMC 4N (5nm)	TSMC 4N (5nm)	TSMC 4N (5nm)

* 4090 的 AD102-300 核心中每个 SM 单元中有 128 个 CUDA 计算单元，其中 64 个 CUDA 可以计算 FP32 或 INT32，另外 64 个只能计算 INT32。

** 第一项为 Tensor Core 计算 FP8、FP16、BF16、TF32 时的加速频率，第二项为 Tensor Core 计算 FP64 和 CUDA Core 计算 FP32、FP64 时的加速频率。

二、算力

1、CUDA Core 算力

浮点：TFLOPS

整型：TIOPS

取 A100 80GB PCIe 的算力为 100%

	V100 SXM2	V100 PCIe	V100S PCIe	A100 40GB PCIe	A100 80GB PCIe	A100 40GB SXM	A100 80GB SXM	H100 SXM5	H100 PCIe	4090
FP32	15.67	14.13	16.35	19.5	19.5	19.5	19.5	66.9	51.2	82.6
FP16	31.33	28.26	32.71	78	78	78	78	133.8	102.4	82.6
FP64	7.834	7.066	8.177	9.7	9.7	9.7	9.7	33.5	25.6	1.29
BF16	NA	NA	NA	39	39	39	39	133.8	102.4	82.6
INT32	15.67	14.13	16.35	19.5	19.5	19.5	19.5	33.5	25.6	41.3

	V100 SXM2	V100 PCIe	V100S PCIe	A100 40GB PCIe	A100 80GB PCIe	A100 40GB SXM	A100 80GB SXM	H100 SXM5	H100 PCIe	4090
FP32	80.4%	72.5%	83.8%	100%	100%	100%	100%	343%	263%	424%
FP16	40.2%	36.2.%	41.9%	100%	100%	100%	100%	172%	131%	106%
FP64	80.4%	72.5%	83.8%	100%	100%	100%	100%	343%	263%	13.3%
BF16	NA	NA	NA	100%	100%	100%	100%	343%	263%	212%
INT32	80.4%	72.5%	83.8%	100%	100%	100%	100%	172%	131%	212%

2、Tensor Core 算力

浮点：TFLOPS

整型：TIOPS

稠密/稀疏

取 A100 80GB PCIe 的算力为 100%

	V100 SXM2	V100 PCIe	V100S PCIe	A100 40GB PCIe	A100 80GB PCIe	A100 40GB SXM	A100 80GB SXM	H100 SXM5	H100 PCIe	4090
FP8	NA	NA	NA	NA	NA	NA	NA	1978.9 / 3957.8	1513 / 3026	660.6 / 1321.2
FP16	125	112	130	312 / 624	312 / 624	312 / 624	312 / 624	989.4 / 1978.9	756 / 1513	330.3 / 660.6
BF16	NA	NA	NA	312 / 624	312 / 624	312 / 624	312 / 624	989.4 / 1978.9	756 / 1513	165.2 / 330.4
TF32	NA	NA	NA	156 / 312	156 / 312	156 / 312	156 / 312	494.7 / 989.4	378 / 756	82.6 / 165.2
FP64	NA	NA	NA	19.5	19.5	19.5	19.5	66.9	51.2	NA
INT8	NA	NA	NA	624 / 1248	624 / 1248	624 / 1248	624 / 1248	1978.9 / 3957.8	1513 / 3026	660.6 / 1321.2
INT4	NA	NA	NA	1248 / 2496	1248 / 2496	1248 / 2496	1248 / 2496	3957.8 / 7915.6	3026 / 6052	1321.2 / 2642.4
Binary	NA	NA	NA	4992	4992	4992	4992	NA	NA	NA

	V100 SXM2	V100 PCIe	V100S PCIe	A100 40GB PCIe	A100 80GB PCIe	A100 40GB SXM	A100 80GB SXM	H100 SXM5	H100 PCIe	4090
FP8	NA	NA	NA	NA	NA	NA	NA	NA	NA	NA
FP16	40.1%	35.9%	41.7%	100%	100%	100%	100%	317%	242%	106%
BF16	NA	NA	NA	100%	100%	100%	100%	317%	242%	52.9%
TF32	NA	NA	NA	100%	100%	100%	100%	317%	242%	52.9%
FP64	NA	NA	NA	100%	100%	100%	100%	343%	263%	NA
INT8	NA	NA	NA	100%	100%	100%	100%	317%	242%	106%
INT4	NA	NA	NA	100%	100%	100%	100%	317%	242%	106%
Binary	NA	NA	NA	100%	100%	100%	100%	NA	NA	NA