Leaked report: Huawei Ascend 950DT hits 86% of H200 throughput on GLM-5.2 at 0.28 kWh/M tokens
teortaxesTex · x · 2026-11-18
A leaked China Mobile test report reveals real-world numbers for Huawei's 64-card Atlas 950 PoD. Despite lower paper TFLOPS than Ascend 910C, the 950DT delivers 1,951 tokens/s per card on GLM-5.2 and 1,451 on DeepSeek-V4-Pro — 4x faster than 910C and roughly 86% of Nvidia H200's throughput on GLM-5.2, at just 0.28 kWh per million tokens. Economically, serving GLM-5.2 yields an 18.6% profit margin while running DeepSeek-V4-Pro at current API prices loses 34%. Commentator teortaxesTex judges the 950DT roughly H100-class, viable for training and some RL but not economical inference or long agentic RL, and asks how many units will ship by end-2026.
Related event: Leaked China Mobile Report: Ascend 950DT Hits 86% of H200 Throughput(3 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