Shows DeepSeek now ships Ascend-native kernels (TileLang, DeepGEMM, FlashMLA) built on Huawei hardware, evidence that China's frontier AI stack is becoming independent of Nvidia, which changes assumptions about future compute access.
Key takeaways · AI-distilled
SemiAnalysis ran stock torch.matmul on an Ascend 950PR across 1,055 BF16 GEMM shapes: Huawei's best kernels reached 99.7% of the attainable ceiling, though the chip's absolute performance sits between A100 and MI300X.
Ascend's design is changing DeepSeek's code: FlashMLA now stores quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition → KV values next to their scaling factors to cut memory-transfer requests, and attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → keeps its running accumulator in Vector UB because the L0C buffer cannot rescale in place.
Of the six DeepSeek repos that gained Ascend support, only TileLang was usable at writing; DeepSeek's code requires CANN 9.2.0, then unreleased. TileLang's Ascend backend is merged upstream, and its AutoSchedule cut single-operation time by up to 40% at small sizes.
DeepEP-Ascend keeps DeepEP V2.5's Python API over URMA and HCCL. A DeepEP maintainer declined it under deepseek-ai in July 2025, but it now ships under DeepSeek's organization, and its docs flag HBM contention that Huawei plans to ease in firmware.
The authors caution the work fits DeepSeek's own setup best: optimizations target supernodes, and as of vLLM-Ascend 0.26.0rc1 x86 hosts lack DeepGEMM support and hit hardcoded ARM headers that block startup.
Terms in this piece · Glossary
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
