The South China Morning Post reported on 22 October that semiconductor research firm SemiAnalysis sees Huawei and DeepSeek narrowing Nvidia’s AI-chip software gap after DeepSeek released six modules for Huawei’s Ascend platform.
One module brings TileLang, a language for writing code that runs computing tasks on chips, to Ascend. AI Affairs previously reported on DeepSeek’s partnership with Huawei to develop Ascend programming tools.
SemiAnalysis expects DeepSeek to rely primarily on Huawei’s technology stack by 2027 or 2028. Huawei engineers’ best Ascend kernels reached 99.7 per cent of their theoretical efficiency ceiling, the research firm said. Huawei has announced inference support for DeepSeek V4 and V4.1 from their release, after DeepSeek gave it early access to the models.
Huawei’s Ascend 950PR, unveiled in March, trails Nvidia’s H100 in absolute performance, SemiAnalysis said. Most optimisation has targeted large supernodes, while developers using Ascend chips with standard x86 host processors still face compatibility problems, according to the report.
