DeepSeek open-sourced core software tools for Huawei Ascend chips. The tools include TileLang, DeepGEMM, DeepEP, FlashMLA, TileKernels and DeepSelect. TileLang is a high-level language for writing GPU programs. DeepGEMM handles matrix operations for large models. DeepEP manages communication between many chips during training. FlashMLA and others accelerate attention and data tasks. Performance in key tests now approaches Ascend hardware limits. DeepSeek used TileLang for most operators in its V4 model training. Huawei collaborated closely on the development. Both firms are jointly building a 128-card supernode using Ascend 950. The work aims to reduce the cost of moving AI workloads from NVIDIA to Chinese hardware.
What changed
Chinese AI developers relied on NVIDIA CUDA and its ecosystem for training frontier models.
What it unlocks
Training frontier AI models on Huawei Ascend chips with a mature software stack.
What you need to act on it
- Huawei Ascend hardware
- adoption of DeepSeek's open-sourced tools
Sources