The Pulse
DeepSeek Opens Huawei Ascend Programming Tools
DeepSeek said it is partnering with Huawei to develop software for Ascend AI chips and has released an Ascend-focused version of DeepGEMM. The release describes the software’s supported operations, optimization techniques and performance me
AI.info Team · Written with AI ·
“To build a new generation of independent, self-controlled GPU software ecosystem, the first priority is establishing a high-level language that is universal, easy to program, and still capable of reaching the hardware's full performance potential,”
DeepSeek, in a post on its official WeChat account
DeepSeek said on Wednesday, September 30, that it had partnered with Huawei Technologies to develop programming tools for Huawei’s Ascend AI chips. The company said it was open-sourcing programming infrastructure for the platform, including compute and communication libraries.
Tools for the Ascend platform
DeepSeek said Huawei provided full support in developing the programming infrastructure. The collaboration is focused on software for Huawei’s Ascend processors, rather than a new DeepSeek model or a new chip.
DeepSeek also highlighted TileLang, a high-level open-source programming language for AI chips. It said the language is intended to improve development efficiency and simplify code logic. The company described TileLang’s programming model as simpler than Nvidia’s CUDA, an important comparison for developers considering software for processors outside Nvidia’s ecosystem.
The GitHub repository DeepSeek published for DeepGEMM Ascend gives a more specific picture of the software release. It describes the project as a port of DeepGEMM to Huawei’s Ascend platform, designed to remain API-compatible with DeepGEMM. The repository says it supports BF16, FP8 and FP4 general matrix multiplication, as well as MQA logits and MegaMoE.
According to the repository, DeepGEMM Ascend uses an abstraction over Ascend matrix multiply-add operations to hide details such as memory layouts, alignment requirements and address calculations. It also describes Ascend-specific optimizations, including sparse data loading and coroutine-based pipelining. The project’s README says the initial release, dated September 30, supports Ascend 950 devices.
A 128-chip system
DeepSeek said the two companies jointly advanced a “supernode” solution based on 128 Ascend 950 chips, with work to optimize both computation and communication across the system. The announcement came two weeks after Huawei unveiled its next generation of AI processors and supernode computing systems. Huawei said it expected its AI systems to be widely used for model training in 2027.
The repository also describes performance tests on an Ascend 950DT system, using the CANN 9.20 software toolkit. Its reported dense matrix-multiplication results include utilization figures of up to 99.8% of the listed hardware limit for tested cases. Those are repository-reported results for specified workloads; they are not a direct comparison with Nvidia chips.
What the release establishes
The open-source release and the collaboration announcement document work on software and a multi-chip system, but do not establish that DeepSeek has stopped using Nvidia hardware.
For now, the public details are concrete: DeepSeek has published an Ascend-focused software project, described its capabilities and listed performance measurements for particular tests. Whether those tools will make Ascend systems easier to use for a wider range of workloads will depend on developers’ experience with them and the systems themselves.