News
Public · Published
Huawei Ascend 950PR Matrix Computing Speed Reaches 43% of Nvidia H100
Huawei's Ascend 950PR reached matrix-computing speeds equal to 43% of Nvidia's H100 in tests by semiconductor research firm SemiAnalysis, which tested 1,055 matrix operations commonly used in large AI models.
Published:
Updated:
What happened
Huawei’s Ascend 950PR reached matrix-computing speeds equal to 43% of Nvidia’s H100 in tests by semiconductor research firm SemiAnalysis, which tested 1,055 matrix operations commonly used in large AI models.
Confirmed
Global impact / market context
The comparison gives chip buyers a clearer sense of how far Huawei’s hardware can carry AI workloads, which affects purchasing decisions and Nvidia’s competitive position, while DeepSeek’s software work shows that efficiency gains can come from code as well as silicon.
Analyst inference
The 950PR reached only 21% of Nvidia’s newer B300, so Huawei still trails the leading chip, and DeepSeek’s tools target data transfers and waiting times, meaning lower costs per task could shift demand toward cheaper hardware.
Analyst inference
What to watch
- SemiAnalysis found automated scheduling cut the time for some small computing tasks by up to about 40%, showing software tuning can raise effective chip output without new hardware. Confirmed
- Huawei plans further firmware improvements, which means updates to low-level chip software; those changes could raise the 950PR’s attainable performance and lower the cost of running AI models. Proposed
- DeepSeek’s porting of DeepEP, which handles communication among mixture-of-experts models, may ease memory bandwidth competition, that is the shared supply of fast memory, but results are not yet proven. Analyst inference