DeepSeek open-sources its training stack for Huawei's Ascend chips
TL;DR
- On 30 September DeepSeek released open-source software for Huawei's Ascend 950 chips: DeepGEMM-Ascend, DeepEP-Ascend, TileKernels with an Ascend backend, and Ascend versions of TileLang, FlashMLA and DeepSelect, by Dataquest India's count.
- DeepSeek says the first job in building an independent GPU software ecosystem is a high-level language that is easy to program and still reaches the hardware's full performance, and that TileLang is simpler than Nvidia's CUDA.
- DeepGEMM-Ascend reports up to 99.8% hardware utilisation on dense matrix multiplies on an Ascend 950DT. The numbers are DeepSeek's own.
Read the full story
Sign in with your email to read AI News. It’s free.