SemiAnalysis deep-dive: Huawei Ascend 950PR hits 86% median MFU, beating all Nvidia chips, but only A100-tier compute
teortaxesTex · x · 2026-10-22
SemiAnalysis got a Huawei Ascend 950PR (950PR9579) and published microbenchmarks:
- Specs: 600W, 325 TFLOPS median BF16 — 44% of H100, 21% of B300, 2x worse FLOPS/W than Hopper; thoroughly A100-tier.
- Memory: effective bandwidth plateaus at 1.4 TB/s vs 3 TB/s for H100 SXM5, though bandwidth utilization edges out H100 at small message sizes.
- Kernel quality: on compute-bound shapes the 950PR's median MFU is 86%, beating every Nvidia chip; Huawei's kernels reach 99.7% of roofline, rivaling DeepSeek's DeepGEMM. Memory-bound: 41.5% vs H100's 47%.
- Huawei–DeepSeek symbiosis: DeepSeek released Ascend support for TileLang, DeepGEMM, DeepEP, FlashMLA and more — Huawei supplies the stack, DeepSeek the workloads. The piece expects DeepSeek to run primarily on Huawei's stack by 2027–2028.
More from Infra
- Flint 40B hits 128 tok/s on M4 Pro via Kapitsa, 1.9x faster than llama.cpp — DotaMate · 2026-10-22
- kneebench: open-source LLM benchmark that tells you what the numbers mean and which vLLM flag to tweak — s1lv3rj1nx · 2026-10-22
- Open tool computes API vs. RunPod self-hosting breakeven across 450+ models, with VRAM sizing — zackku · 2026-10-22
- Railway breach timeline: crypto theft began 2 hours before users were told to rotate keys — banteg · 2026-10-22
- NVIDIA shows off DGX Spark clusters, Vera Rubin NVL72 trays at GTC Berlin — NVIDIAAI · 2026-10-22
- Court Filing: Meta Told Board It Needs to Raise Hundreds of Billions Starting Fall 2026 — firstadopter · 2026-10-22