The Hangzhou-based AI developer released Ascend versions of several key components on September 30, including its DeepGEMM matrix-multiplication library and DeepEP communications software. The projects are designed to work with Huawei’s Ascend processors while keeping public interfaces aligned with versions built around Nvidia hardware, reducing the amount of code developers may need to rewrite when moving workloads.
That portability targets a critical part of Nvidia’s competitive advantage. Nvidia’s CUDA software platform has become deeply embedded in AI development, giving programmers mature libraries, tools and optimisation techniques built over years. Competing chips therefore have to overcome not only differences in raw computing performance but also the cost and complexity of shifting software away from CUDA.
DeepSeek’s DeepJIT project takes a further step by providing a shared runtime interface for Nvidia CUDA graphics processors and Huawei Ascend neural-processing units. It handles compilation, caching, loading and launching of specialised AI kernels while retaining backend-specific code where necessary. The approach does not make the two architectures interchangeable, but it can reduce friction for developers maintaining software for both.
DeepGEMM-Ascend is fully API-compatible with DeepSeek’s original DeepGEMM package and supports BF16, FP8 and FP4 calculations used in demanding AI workloads. DeepSeek says the library uses Ascend-specific optimisation methods to approach the hardware’s performance limits while shielding programmers from lower-level complexities including memory layouts, alignment constraints and address calculations.
Its DeepEP-Ascend library similarly aligns its public buffer interfaces with the Nvidia version while using Huawei’s communications technologies underneath. DeepSeek has also added Ascend support to DeepSelect, a specialised component used for sparse-attention and sampling operations, and released Ascend sparse-attention kernels for its model infrastructure.
Huawei, meanwhile, has been opening more of the software surrounding Ascend. Its CANN platform — the company’s counterpart to the programming infrastructure required to exploit AI accelerators — supports frameworks including PyTorch and works with open-source projects such as Triton, vLLM and veRL.
Huawei said last month that CANN had moved to sustained community-driven open-source development, with external programmers accounting for 61 per cent of its CANN developers. The company said the community had more than 5,200 monthly active developers, while more than 40 models had been natively pre-trained using Ascend and CANN.
The combination matters because China’s challenge to Nvidia has increasingly shifted from producing an alternative accelerator to assembling a usable full computing stack. Hardware can win orders only if model developers can train and run workloads efficiently without sacrificing too much engineering time or performance.
DeepSeek’s software also gives Huawei an influential independent developer helping test and improve that stack. The AI laboratory became prominent for building capable models with an emphasis on computational efficiency, and its infrastructure code is closely watched by developers seeking lower-cost approaches to training and inference.
The effort comes as US export controls continue to restrict China’s access to Nvidia’s most advanced processors, creating stronger incentives for technology groups to adopt domestic alternatives. Those restrictions have helped expand the strategic importance of Huawei’s Ascend line, even though China still faces constraints in advanced semiconductor manufacturing.
Nvidia nevertheless retains formidable advantages. CUDA has a vast installed developer base and an extensive collection of mature libraries, while Nvidia’s leading accelerators remain central to major AI training systems outside China. Compatibility layers and open-source libraries also require sustained testing, documentation and optimisation before they can match the reliability of an established platform across diverse workloads.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
