Technology & Science
DeepSeek opens software tools for Huawei's Ascend AI chips
New DeepSeek repositories include matrix-compute and communications libraries for Ascend processors. The announcement advances a domestic software ecosystem, but no public benchmark or deployment figures show how much AI work can move off Nvidia hardware.
Executive summary
- What changed: DeepSeek announced an open-source software push for Huawei's Ascend AI chips. Reuters reported that the release includes TileLang and related compute and communication libraries; the companies described the work as a partnership to develop Ascend programming tools.[1][2]
- What is public: DeepSeek's GitHub organization now includes Ascend-specific matrix-compute and communication projects, while its TileKernels repository documents optimized kernels. Repository availability makes code inspectable; it does not show that the stack has been adopted at scale or that its performance matches Nvidia's tools.[3][4][5][1]
- Why it matters: AI accelerators need software that lets developers write, optimize and connect workloads as well as chips. A wider Ascend toolchain could make Huawei hardware easier for Chinese AI developers to use, but the announcement gives no workload migration, benchmark or cost figures.[1][2][3][4]
DeepSeek said it is working with Huawei Technologies on programming tools for Huawei's Ascend processors and is releasing Ascend-platform infrastructure as open-source software. Reuters reported that the release includes TileLang, a higher-level programming language, alongside compute and communication libraries. The South China Morning Post reported a six-module release and identified Ascend support in the TileLang project.[1][2]
The public code gives a more concrete view of the effort. DeepSeek's GitHub organization lists DeepGEMM-Ascend, a matrix-multiplication library for Huawei neural-processing units, and DeepEP-Ascend, a communication library for machine-learning training and inference. Its TileKernels project describes optimized kernels written with TileLang. These are software components for making accelerator workloads run; they are not new chip announcements.[3][4][5]
Reuters also reported that the companies jointly advanced a supernode design using 128 Ascend 950 chips, with the software work aimed at both computation and communication. The announcement did not include independent benchmarks, operating costs or comparisons with a similarly sized Nvidia system, so the design is evidence of engineering work, not proof of equivalent performance.[1]
The distinction matters because a programming layer can lower one barrier to using different hardware without removing the others. The repositories document code and intended uses, but the public announcement does not provide independent performance tests, production deployments, customer counts or evidence that the tools can replace Nvidia's CUDA ecosystem across a broad range of workloads.[3][4][5][1][2]
Reuters framed the partnership as another step by Chinese technology companies seeking alternatives to Nvidia's software and hardware ecosystem. The initial evidence is a set of open-source tools and repositories, not disclosed chip orders or a quantified migration of DeepSeek's training and inference workloads. DeepSeek and Huawei did not publish a financial value or timetable for the collaboration in the announcement reported by Reuters.[1][2][3][4]
That leaves practical questions open: how well the tools perform on Ascend hardware, how much engineering is needed to port existing workloads, and whether outside developers adopt the projects. Open-source code can invite review and contributions, but publication alone does not establish reliability, compatibility or commercial use at scale.[3][4][5][2]