英伟达和华为的性能比较

What’s Known / Projected for Nvidia

Some of Nvidia’s relevant systems (public or projected):

• H100: current workhorse GPU for many large models, high FP16 / FP8 / mixed precision performance, good interconnect (NVLink), strong ecosystem (CUDA, cuDNN, etc.).
• GB200 / NVL72: A system using Blackwell-based GB200 GPUs; Nvidia is pushing forward NVL / Rubin / “Vera Rubin” scale platforms.
• Nvidia plans NVL144 in 2026, NVL576 in 2027, etc.  

From Huawei’s claims during their announcements:

• Huawei claims Atlas 950 will have 6.7 more compute power, 15 more memory capacity, ~16.3 PB/s interconnect bandwidth compared to NVL144.  
• They also claim Atlas 950 will outperform systems Nvidia plans in 2027 (e.g. NVL576) across “all fronts.”