Huawei’s Ascend 950DT could beat Nvidia B300 on tokens per watt, analysis says
teortaxesTex · x · 2026-07-21
A quoted analysis argues that Huawei’s next-gen Ascend 950DT may beat Nvidia’s B300 on tokens-per-watt, while the current CloudMatrix 384 generation is already at parity with the token rates API providers target.
The comparison is based on announced specs, assumed power numbers, and calibration from Huawei’s published DeepSeek-R1 serving results, with K3 geometry inferred ahead of the 7/27 weight drop.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27