DeepSeek and Huawei co-develop Ascend software, with benchmarks and remaining gaps
AIDeepSeek released Ascend support for six of its open-source repositories, including TileLang, DeepGEMM, DeepEP, and FlashMLA, in a trade where Huawei supplies hardware and DeepSeek supplies models and workloads. SemiAnalysis benchmarks the Ascend 950PR, finding its best BF16 GEMM kernels reach 99.7% of the ceiling, while its memory bandwidth reaches 1.6384 TB/s.
AIWhy it matters.
The article shows how DeepSeek's kernel work and Huawei's chip stack depend on each other, with benchmarks and an execution profile showing where the gains and remaining gaps are.
Source details
SemiAnalysis · Bryan Shan · newsletter.semianalysis.comPublished · added here
Original title: The Huawei-DeepSeek Symbiosis