DeepSeek broadens Huawei software challenge to Nvidia
DeepSeek has expanded support for Huawei’s Ascend artificial-intelligence processors by releasing core programming libraries designed to preserve familiar developer workflows across rival hardware, strengthening China’s effort to loosen Nvidia’s grip on AI computing software. The Hangzhou-based AI developer released Ascend versions of several key components on September 30, including its DeepGEMM…
DeepSeek has extended its support for Huawei's Ascend artificial-intelligence processors by introducing core programming libraries that maintain familiar developer workflows across different hardware. This move bolsters China's strategy to reduce Nvidia's dominance in AI computing software. On September 30, DeepSeek unveiled Ascend versions of significant components, such as its DeepGEMM matrix-multiplication library and DeepEP communications software.
These projects are designed to operate with Huawei's Ascend processors while maintaining compatibility with Nvidia hardware interfaces, thereby minimizing code modifications when transferring workloads. This portability addresses a crucial aspect of Nvidia's competitive edge. CUDA software, developed by Nvidia, has become a cornerstone of AI development, providing programmers with mature libraries, tools, and optimization techniques honed over years.
Consequently, competing chips must overcome not only differences in raw computing power but also the financial and technical challenges of transitioning software away from CUDA. To further enhance compatibility, DeepSeek introduced DeepJIT, a shared runtime interface for Nvidia CUDA GPUs and Huawei's Ascend neural-processing units.
This interface manages compilation, caching, loading, and launching of specialized AI kernels while preserving backend-specific code as needed. Although this approach does not render the two architectures interchangeable, it can lessen the burden 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, which are crucial for demanding AI workloads.
DeepSeek claims that DeepGEMM-Ascend utilizes Ascend-specific optimization techniques to maximize hardware performance while abstracting lower-level complexities such as memory layouts, alignment constraints, and address calculations. Similarly, DeepSeek's DeepEP-Ascend library aligns its public buffer interfaces with Nvidia's counterpart while leveraging Huawei's communications technologies beneath the surface.
In addition, DeepSeek has incorporated Ascend support into DeepSelect, a specialized component for sparse-attention and sampling operations, and released Ascend sparse-attention kernels for its model infrastructure. Meanwhile, Huawei has been progressively opening more of the software surrounding Ascend. Its CANN (China Artificial Intelligence Neural Network) platform, an alternative to the programming infrastructure needed to harness AI accelerators, supports frameworks like PyTorch and integrates with open-source projects such as Triton, vLLM, and veRL.
Huawei announced last month that CANN has transitioned to community-driven open-source development, with external programmers constituting 61% of its CANN contributors. The platform now boasts over 5,200 monthly active developers, and more than 40 models have been pre-trained using Ascend and CANN. This development is significant as China's attempt to challenge Nvidia has evolved from creating an alternative accelerator to establishing a comprehensive computing stack.
Hardware can only secure orders if model developers can train and run workloads efficiently without incurring excessive engineering time or sacrificing performance. DeepSeek's software provides Huawei with a prominent independent developer who aids in testing and improving the stack. This effort is particularly crucial given the US export controls that have further limited China's access to Nvidia's most advanced processors, thereby strengthening the case for domestic alternatives.
Although China continues to face challenges in advanced semiconductor manufacturing, Huawei's Ascend line has gained strategic importance. Despite Nvidia's formidable advantages, including a vast installed developer base and a rich collection of mature libraries, compatibility layers and open-source libraries require continuous testing, documentation, and optimization to match the reliability of an established platform across diverse workloads.
Written by urgent.news from Arabian Post's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.