China Mobile test: Ascend 950DT inference hits ~86% of H200, wins on cost
basedjensen · x · 2026-11-17
Citing tphuang: China Mobile benchmarked Huawei's Ascend 950DT against the 910 series and H200 for GLM inference. Results: 950DT delivered 11x the throughput of 910B and 3x that of 910C, only 14% below H200 — despite lower TFLOPS and 1.3x the cost of 910C.
Key takeaway: system-level design (64-card cabinet, high-speed interconnect, larger and higher-bandwidth HBM) matters more than per-chip compute specs for inference. Factoring in rental costs, 950DT economically beats H200 — which is why Chinese AI labs want B300 rather than H200. The retweeter quips that sea data centers pivoting to Ascend would enrage the export-control crowd.
More from Infra
- NVIDIA open-sources cuPotential, speeding molecular dynamics up to ~4x on GPUs — AllThingsApx · 2026-11-17
- PiPNN removes the search step from graph index builds, cutting billion-vector build times from days to minutes — qdrant_engine · 2026-11-17
- tinygrad.js runs the real tinygrad in your browser on your GPU via WebGPU — CatAstro_Piyush · 2026-11-17
- Data Center Backlash Grows Loud, With Historical Echoes of 'A Nuisance Has No Rights' — justin_hart · 2026-11-17
- Constrained Decoding for Function Calling Cuts Endpoint Throughput by 17%-90% — remilouf · 2026-11-17
- Terray taps NVIDIA CUDA kernels to speed drug-design model TerraBind up to 7x — AllThingsApx · 2026-11-17