DeepSeek open-sources infrastructure components for the Huawei Ascend platform.

ChainCatcher report: AI company DeepSeek has officially open-sourced infrastructure components tailored for Huawei Ascend computing platforms, including the TileLang high-level language compiler, compute libraries, and distributed communication libraries—each corresponding directly to previously open-sourced components for NVIDIA platforms. TileLang aims to provide a universal, easier-to-program high-level language that achieves maximum hardware performance, improving development efficiency and simplifying logic compared to CUDA, while its programming model fully leverages chip-specific features. The TileLang roadmap was first validated on NVIDIA platforms, where it already implements the majority of operators used in DeepSeek V4 model training. The newly open-sourced Ascend version encapsulates Ascend C low-level instructions, offering a high-level programming interface without sacrificing hardware performance. Every TileLang operator used in DeepSeek’s training now has a corresponding high-performance implementation on Ascend. Simultaneously open-sourced components include DeepGEMM (general matrix operations), DeepEP (large-scale cross-device communication), TileKernels (standard vector computations and memory access), FlashMLA (long-context sparse attention), and DeepSelect (data filtering). DeepSeek states that key benchmarks show computational and communication performance approaching hardware limits; during development, Huawei’s team provided support, and both parties collaborated to advance a 128-card super-node solution based on Ascend 950, with deep optimizations in computation and communication.