Chinese artificial intelligence developer DeepSeek has partnered with Huawei Technologies to develop and open-source programming infrastructure for Huawei’s Ascend chips, expanding the software available to developers building AI workloads on Chinese-designed accelerators.
Reuters reported on September 30 that DeepSeek disclosed the work through its official WeChat account. The companies said the effort includes compute and communication libraries for Ascend, as well as work on a supernode system built around 128 Ascend 950 chips.
Open-source components target Ascend development
The collaboration addresses a part of the AI computing stack that is less visible than the processor itself but critical to adoption. Developers need compilers, optimized mathematical kernels and communication libraries to translate model workloads into operations that can run efficiently across many accelerators.
DeepSeek said Huawei provided full support for the programming infrastructure. The partners worked on both computation and communication for the 128-chip supernode, according to Reuters. The announcement did not provide independent performance results or disclose production deployments using the full stack, so comparisons with established accelerator platforms remain unverified.
DeepSeek also highlighted TileLang, an open-source high-level language intended to simplify the development of performance-sensitive AI kernels. The TileLang project describes itself as a domain-specific language for high-performance kernels across GPUs, CPUs and other accelerators. A separate Ascend adapter repository targets Huawei’s neural processing units.
The companies’ stated goal is to give developers a programming model that is easier to use while still exposing enough of the underlying hardware to reach high performance. That is an important tradeoff: lower-level tools can deliver efficiency but require specialized engineering, while higher-level abstractions can reduce development time at the risk of leaving hardware capacity unused.
Software is central to the accelerator contest
Nvidia’s CUDA platform combines programming tools, optimized libraries and a large developer ecosystem around its GPUs. Hardware alternatives therefore compete not only on chip specifications, but also on whether developers can move existing models, tune kernels and operate large clusters without extensive rewrites.
DeepSeek’s release adds application-level experience to Huawei’s effort to build a fuller Ascend ecosystem. The AI developer has already worked with Ascend hardware on model deployment and optimization, giving it practical knowledge of the bottlenecks faced by model teams. Open-sourcing components may also invite external developers to inspect, adapt and improve the tools, although adoption will depend on documentation, stability, hardware access and compatibility with commonly used AI frameworks.
The move follows Huawei’s September disclosure of new AI processors and large-scale computing systems. TNGlobal previously reported that Huawei is advancing its Ascend roadmap and a system architecture designed to connect very large numbers of processors. The new DeepSeek collaboration focuses on the software needed to turn that hardware capacity into usable model-training and inference performance.
Huawei unveils million-processor architecture as it accelerates AI chip roadmap