SemiAnalysis 实测华为 Ascend 950PR：性能约为 H100 的 44%
SemiAnalysis 真拿到了华为 Ascend 950PR 实测：算力只有 H100 的 44%，但算子效率比 Nvidia 还高，数据很细。
SemiAnalysis 拿到华为 Ascend 950PR（950PR_9579）进行实测。该芯片功耗 600W，BF16 算力中位数为 325 TFLOPS，约为 H100 的 44%、B300 的 21%，FLOPS/W 比 Hopper 低 2 倍，实际达到 A100 档位。有效内存带宽约 1.4 TB/s，低于 H100 SXM5 的 3TB/s。不过在计算密集型负载下，950PR 的 MFU 中位数达 86%，超过所有 Nvidia 芯片，内存密集型场景 MFU 为 41.5%，低于 H100 的 47%。
3 weeks later I have my answer: SemiAnalysis somehow got a Huawei Ascend 950PR, specifically 950PR_9579.
Interesting details: - 600W, 325 TFLOPS BF16 median. 44% of H100, 21% of B300. 2x worse than Hopper on FLOPS/W. A100 tier really - Effective memory bandwidth plateaus at about 1.4 TB/s vs 3TB/s for H100 SXM5 - on compute-bound shapes, 950PR’s median MFU is 86%, beating all Nvidia. Huawei's kernels are as good as DS's. If memory-bound, 41.5% (vs 47% for H100)
it's a thoroughly entry-level chip, yeah