HuaWei、NVIDIA 数据中心 AI 算力对比

HuaWei & NVIDIA 之 AI 算力对比

参数/型号	A800	A100	Huawei Ascend 910B	H800	H100	H200
Year	2022	2020	2023	2022	2022	2024
Manufacturing	7nm	7nm	7+nm	4nm	4nm	4nm
Architecture	Ampere	Ampere	HUAWEI Da Vinci	Hopper	Hopper	Hopper
Max Power	300/400 W	300/400 W	400 W	350/700 W	350/700 W	700W
GPU Mem	80G HBM2e	80G HBM2e	64G HBM2e	80G HBM3	80G HBM3	141GB HBM3e
GPU Mem BW	1935/2039 GB/s			2/3.35 TB/s		4.8 TB/s
GPU		PCIe Gen4			PCIe Gen5	PCIe Gen5
Interconnect (one-to-one max bw)	NVLINK 400GB/s	NVLINK 600GB/s	HCCS 56GB/s	NVLINK 400GB/s	NVLINK 900GB/s	NVLINK 900GB/s
GPU		PCIe Gen4			PCIe Gen5	PCIe Gen5
Interconnect (one-to-many total bw)	NVLINK 400GB/s	NVLINK 600GB/s	HCCS 392GB/s	NVLINK 400GB/s	NVLINK 900GB/s	NVLINK 900GB/s
FP32 TFLOPS		19.5			51	67*
TF32 TFLOPS		156	312*		756	989*
BF16 TFLOPS		156	312*		1513	1979*
FP16 TFLOPS		312	624*	328	1513	1979*
FP8 TFLOPS	NOT support	NOT support			3026	3958*
FP6 TFLOPS	NOT support	NOT support
FP4 TFLOPS	NOT support	NOT support
INT8 TFLOPS		624	1248*	648	3026	3958*

HuaWei Ascend 910B 是 HuaWei 于 2023 年推出的高性能 AI 处理器芯片，其对标产品为 Nvidia A100/A800

---

NVIDIA 中国特供版

参数	HGX H20	L20 PCIe (Ada Lovelace)	L2 PCIe (Ada Lovelace)
GPU Architecture	NVIDIA Hopper	NVIDIA Ada Lovelace	NVIDIA Ada Lovelace
GPU Memory	96 GB HBM3	48 GB GDDR6 w/ ECC	24 GB GDDR6 w/ ECC
GPU Memory Bandwidth	4.0 TB/s	864 GB/s	300 GB/s
INT8 / FP8 Tensor Core*	296 TFLOPS	239 TFLOPS	193 TFLOPS
BF16 / FP16 Tensor Core*	148 TFLOPS	119.5 TFLOPS	96.5 TFLOPS
TF32 Tensor Core*	74 TFLOPS	59.8 TFLOPS	48.3 TFLOPS
FP32	44 TFLOPS	59.8 TFLOPS	24.1 TFLOPS
FP64	1 TFLOPS	N/A	N/A
RT Core	N/A	Yes	Yes
MIG	Up to 7 MIG	N/A	N/A
L2 Cache	60 MB	96 MB	36 MB
Media Engine	7 NVDEC 7 NVJPEG	3 NVENC (+AV1) 3 NVDEC 4 NVJPEG	2 NVENC (+AV1) 4 NVDEC 4 NVJPEG
Power	400 W	275W	TBD
Form Factor	8-way HGX	2-slot FHFL	1-slot LP
Interconnect	Pcie Gen5 x16: 128 GB/s ，NVLink: 900GB/s	PCIe Gen4 x16: 64 GB/s	PCIe Gen4 x16: 64 GB/s
Availability	PS: Nov 2023 MP: Dec 2023	PS: Nov 2023 MP: Dec 2023	PS: Dec 2023 MP: Jan 2024

---

NVIDIA Blackwell 架构

NVIDIA GB200 Superchip

为每个 Blackwell GPU 提供高达 192 GB 的 HBM3e 内存

NVIDIA GB200 Superchip

• 组成：由 2 个 NVIDIA Blackwell GPU 和 1 个 NVIDIA Grace CPU 组成，通过 NVIDIA NVLink-C2C 互连技术连接
• 功耗：每个 Blackwell GPU 的满配 TDP 达到 1200W，整个超级芯片的 TDP 达到 2700W
• 第五代 NVIDIA NVLink：提供 1.8TB/s 的 GPU 间互联带宽
• 内存：GB200 Superchip 每个超级芯片配备 864GB 内存（480GB LPDDR5x 和 384GB HBM3e）

算力指标

GPU Spec	GB200 Grace Blackwell Superchip
Configuration	1 Grace CPU : 2 Blackwell GPUs
FP4 Tensor Core Dense/Sparse	20 / 40 petaFLOPS