GPU性能比较

2025-04-09 技术学习

GPU 性能比较

A100/A800/H100/H800/910B/H200/H20/L20/4090比较

	A800 (PCIe/SXM）	A100 (PCIe/SXM)	H800 (PCIe/SXM)	H100 (PCIe/SXM)	H200 (PCIe/SXM)	H20	L20 (PCIe)	4090D	5090D
Year	2022	2022	2023	2024	2024	2023	2023	2022	2024
Manufacturing	7nm	5nm	7nm	5nm	5nm	5nm	5nm	5nm	5nm
Architecture	Ampere	Ampere	Hopper (阉割)	Hopper	Hopper	Hopper	Ada Lovelace	Ada Lovelace	40MB
Max Power	300/400 W	300/400 W	300-350/700W	350-400/700 W	600/700W	400W	350W	425W	575W
CUDA	8.0	8.0	9.0	9.0	9.0	-	8.9	9.0	12.0
GPU Mem	80G HBM2e	80G HBM2e	80G HBM3	64G/94G/80G HBM3	141GB HBM3e	96G HBM3	48G GDDR6	24GB GDDR6X	32 GB GDDR7
GPU Mem BW	2TB/s	1.9/2TB/s	2TB/s	3.9/3.35 TB/s	4.8 TB/s	4TB/s	864GB/s	1008 GB/s	1792 GB/s
GPU Interconnect (one-to-one max bw)	PCIe 4.0：64GB/s NVLink：400GB/s	PCIe Gen4 64GB/s NVLINK 600GB/s	NVLink: 400GB/s PCIe 5.0: 128GB/s	PCIe Gen5 128GB/s NVLINK 600/900GB/s	PCIe Gen5 128GB/s NVLINK 900 GB/s	PCIe Gen5 128GB/s NVLINK 900GB/s	PCIe Gen4 64GB/s	PCIe 4.0 x16 (单向32 GB/s, 双向64 GB/s)	-
GPU Interconnect (one-to-many total bw)	2TB/s	1.9/2TB/s	2TB/s	3.9/3.35 TB/s	4.8 TB/s	4TB/s	864GB/s	1TB/s	-
FP32 TFLOPS	19.5	19.5	51 \| 67	60 \| 67	60 \| 67	44	59.8	-	-
TF32 TFLOPS	156 \| 312	156 \| 312	756 \| 989	989 \| 853	989 \| 853	74	59.8	-	-
BF16 TFLOPS	156 \| 312	312 \| 624	1513 \| 1979	1671 \| 1979	1671 \| 1979	148 \| 148	119 \| 119	-	-
FP16 TFLOPS	156 \| 312	312 \| 624	1513 \| 1979	1671 \| 1979	1671 \| 1979	148 \| 14	119 \| 119	-	-
FP8 TFLOPS	NOT support	NOT support	3026 \| 3958	3341 \| 3958	3341 \| 3958	296 \| 296	-	-	-
INT8 TOPS	NOT support	624 \| 1248	3026 \| 3958	3341 \| 3958	3341 \| 3958	296 \| 296	-	-	-
L1 Cache	192 KB	192 KB	256 KB	256 KB	256 KB	-	128 KB	128 KB	128 KB
L2 Cache	40MB	80MB	50MB	50MB	50MB	60MB	96MB	72MB	96MB