DeepSeek has open-sourced six software components designed for Huawei Ascend AI chips, creating an alternative to the Nvidia software ecosystem. The new toolkit includes TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, which mirror tools previously developed for Nvidia hardware.
TileLang serves as a language for building high-performance computing kernels and now supports Huawei Ascend 950 accelerators. It enables native code generation, automated scheduling, and synchronization. The project aims to establish a highly accessible software ecosystem for various AI chips.
The DeepGEMM Ascend library handles matrix computations with support for BF16, FP8, and FP4 formats, while DeepEP optimizes data exchange between chips for Mixture of Experts architectures. Other components focus on memory operations, long context processing, and sparse attention.
DeepSeek and Huawei collaborated on a supernode system comprising 128 Ascend 950 chips, optimizing both computational tasks and inter-chip communication. The companies plan to continue their joint development of software tools for the Ascend platform.
