Ascend 950 FP8 training decoded: ~217-279 TFLOPS per NPU vs DeepSeek's 385 on H800
teortaxesTex · x · 2026-11-21
Huawei published openPangu FP8 low-precision training results on the Ascend 950 super node, claiming <0.16% loss error, 51% MFU, and 24% speedup over BF16, with plans to open-source the stack. But teortaxesTex breaks down what the numbers actually mean: the 51% MFU is relative to the BF16 peak. The 950DT advertises 547 TFLOPS at the top 36-Cube-Core bin (486 and 425 on lower bins), implying a realistic 217-279 TFLOPS per NPU — versus the 385 TFLOPS DeepSeek achieved on H800s back in 2024. Attached slides show a 1024-card run with 64-way expert parallelism and sequence length 4096.
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