DeepSeek open-sources DeepGEMM-Ascend: MIT-licensed kernels for Huawei Ascend 950
lmoroney · x · 2026-10-03
DeepSeek 于 9 月 30 日开源 DeepGEMM-Ascend，为华为昇腾 950 提供 MIT 协议的矩阵运算内核，作者 Laur ic Moroney 认为这是研究非 CUDA 加速器栈的现实契机。
- 覆盖 BF16、FP8、FP4 GEMM，附带 MQA logits 与 MegaMoE 算子，API 与现有 DeepGEMM 接口兼容
- 稀疏注意力与通信相关组件也在同一波昇腾发布中
- 对开发者的建议：将其当作可审查的内核层，先试点、锁版本、用自己的真实 shape 测量性能
- 对教学者：可布置为作业——盘点服务实际依赖哪些算子，再估算剩余集成成本，而不是等 CUDA 替代品完全成熟
More from Infra
- Read-only Postgres replicas won't stop LLM agents from bloating your primary DB: 4 failure modes — EmetInteractive · 2026-10-03
- Cathie Wood: 90% of global data center debt financing lands in US at ~8% effective tax — rohanpaul_ai · 2026-10-03
- DDR4 + 7900XTX Runs Qwen3-Next at 45-50 tok/s via Strata, Double llama.cpp Speed — EmPips · 2026-10-03
- Musk says growth will far exceed forecasts as energy remains the bottleneck — RachelVT42 · 2026-10-03
- Running Qwen3 27B agent on 16GB VRAM: MTP draft trades context and quality for speed — Remarkable_Air_8383 · 2026-10-03
- DwarfStar 4 (ds4) lets you run DeepSeek V4.1, Qwen and GLM locally — yogthos · 2026-10-03