Chinese AI startup DeepSeek open-sourced on September 30 a comprehensive suite of programming tools designed for Huawei’s Ascend AI accelerators, including the TileLang language as an alternative to Nvidia’s CUDA ecosystem. The toolkit bundles specialized libraries such as DeepGEMM, FlashMLA, TileKernel, DeepSelect, and DeepEP, optimized for Ascend 950 chips and supporting a “supernode” configuration linking 128 accelerators together. The tools are available for free download to lower switching costs for developers who might otherwise default to Nvidia’s ecosystem. This release directly addresses the strategic gap created by US export controls restricting access to advanced Nvidia chips, giving Chinese developers a viable software stack for domestic hardware. DeepSeek also plans to build a data center in Inner Mongolia hosting at least 160,000 Huawei Ascend accelerators.
Source: Read the original article
