SLAI posts 34.22% MFU post-training DeepSeek-V4 on Ascend NPUs

SLAI post-trained DeepSeek-V4 to 34.22% MFU on an Ascend NPU SuperPOD, a 2.93 times improvement over the open source baseline recipe, per the arXiv preprint. Hitting that number on non-Nvidia hardware weakens the assumption, common in Western coverage and baked into export-control policy, that frontier-scale training still requires Nvidia GPUs.

MFU measures how much of a chip's peak compute a training run captures, so 34.22% means SLAI used about a third of the SuperPOD's ceiling. The gains came from optimizing three layers together: model parallelism, computation-communication orchestration, and low-level kernel code tuned to the SuperPOD's architecture. The paper reports training stability held throughout the run for a trillion-parameter MoE model.

A specialized version, DeepSeek-V4-Flash, scored 71.81% zero-shot Pass@1 on Operations Research tasks, 3.98 points ahead of GPT-5.4-Mini on the same benchmark.

SLAI publishing full non-Nvidia training internals, something Western labs rarely do even on GPUs, matters more than whether 34.22% MFU survives outside scrutiny. Watch whether more full training-stack writeups like this one come out of Ascend-based labs than Nvidia-based ones over the next year. That's the sign the training-stack gap is closing faster than the export-control conversation assumes.

As of July 23, most Western AI reporting still assumes frontier-scale training in China depends on Nvidia hardware.