Auditing Huawei Ascend FP8 training: is the '51% MFU' really just 22-25%?
teortaxesTex · x · 2026-11-21
teortaxesTex audits the Ascend FP8 training report's claimed 51% MFU. Key points:
- Definition: '51%' is MXFP8 relative to BF16 peak. The 950DT advertises 547 TFLOPS on the top 36-Cube-Core bin (486/425 on lower bins), so realistically 217-279 TFLOPS per NPU — vs DeepSeek's 385 TFLOPS on H800s in 2024.
- Model caveats: Pangu is 505B total/18B active vs V3's 671/37; its step time also includes the costlier Muon optimizer plus mHC and SWA.
- Recomputation: 33.5M tokens per step (32,768/card), 39.4s with MegaMoE, 830 tokens/s/card → 90 TFLOPS/card, 115-120 with attention — roughly 22-25% of 500 TFLOPS, not 51%.
- Comparison: DeepSeek ran 1,540 tokens/s per H800 on a model with twice the active params; if Table 3-2 matches the headline run, one Ascend card delivers about a quarter of an H800, not 60-65%.
- Caveats: possibly different profiling config, peak-figure accounting, or FLOPs counting; the report gives no absolute tokens/s, so 60-65% should be treated as an upper bound.
Related event: Huawei Ascend 950 FP8 training results spark MFU dispute(2 posts)→
More from Infra
- MLX over Vulkan hits 124 tok/s on RTX 5060 Ti, beating llama.cpp CUDA by 25% — SumitGup · 2026-11-22
- Intruder swapped Prime Intellect's hosted DeepSeek for an abliterated copy for 53 hours — mawcopolow · 2026-11-22
- Token economics model projects $1.14T global API spend by 2030, but only 1% of users pay beyond coding — AccBalanced · 2026-11-22
- Zero-degradation Strata tuning: fixing layer_split lifts decode ~10-15% and halves CPU expert misses — PathfinderTactician · 2026-11-22
- llama.cpp Vulkan SCALE fix PR #32735 delivers 2.3-3.2x speedup on Radeon R9700 — pmttyji · 2026-11-22
- $1400 decade-old V100 32GB runs Qwen 27B: Q4_K_M +48% gen speed, Q8_0 +16% prefill — PuzzleheadedArm3732 · 2026-11-22