- Published on
DeepSeek Open-Sources Infrastructure Components for Huawei’s Ascend Platform, Mapping One-to-One to Nvidia’s Platform
- Authors
- Name
- DeepHorus
DeepSeek announced on the morning of September 30, through its official WeChat public account, that it had officially open-sourced a set of infrastructure components for Huawei's Ascend AI computing platform. The components cover TileLang high-level language compiler tooling, core computing libraries, and distributed communication libraries. The company said that these components correspond one-to-one with projects previously made available for the NVIDIA platform, enabling Ascend hardware to receive more comprehensive high-performance software support. This move is regarded as an important step in DeepSeek's further expansion of its support for domestic computing hardware. The relevant code has been made available to the community through the official GitHub repository.
Large-model training and inference depend not only on GPUs or AI accelerators themselves, but also on compilers, operator libraries, communication frameworks, and higher-level programming tools. NVIDIA's CUDA has long provided a mature toolchain and developer ecosystem, constituting an important advantage at the software level. DeepSeek previously validated TileLang on the NVIDIA platform and used it to implement most of the operators in its V4 series of models. This open-source release amounts to extending that software-stack migration experience to Ascend, reducing the barrier for developers adapting workloads to the domestic platform.
According to the disclosures, the Ascend versions of the components include DeepGEMM for general matrix computation, DeepEP for large-scale inter-device communication, and TileKernels, which provides vector-computation and memory-access operators. FlashMLA focuses on sparse attention and can improve the processing efficiency of long-context tasks, while DeepSelect is used for efficient data selection. DeepSeek stated that every TileLang operator used during its training process currently has a corresponding high-performance implementation on the Ascend platform. The company also said that computation and communication performance in several key tests had already approached the hardware limits, but it did not publish specific test data or standardized benchmarks.
Technically, the Ascend version of TileLang wraps low-level Ascend C instructions and provides programming interfaces through a high-level language, while making every effort to avoid sacrificing hardware performance. DeepSeek said that, compared with CUDA, TileLang offers a simpler programming model, can reduce boilerplate code, and simplify operator logic; its programming model can also be optimized for the characteristics of the chip. Libraries such as DeepGEMM and DeepEP address compute-intensive tasks and multi-card communication bottlenecks, respectively, with the aim of covering the most common high-load scenarios in large-model training.
This adaptation was not a one-sided porting effort. DeepSeek revealed that the Huawei team provided direct support during the development process, and the two sides jointly advanced a 128-card supernode solution based on Ascend 950. The teams also conducted in-depth optimization of computation and communication paths to address bandwidth, latency, and synchronization issues at supernode scale. Public information did not disclose the specific throughput, training efficiency, or comparison results of the 128-card solution against the NVIDIA platform.
For developers, open-sourcing the components means that the Ascend platform has gained more reusable operators and communication tools, so they do not have to redevelop everything from low-level interfaces. For model vendors, a more mature software stack can help shorten the time required to migrate from the NVIDIA platform to Ascend and improve the deployability of large-scale training workloads. However, the actual migration results will still depend on the chip model, framework version, operator coverage, and specific workload. Ecosystem maturity will also determine whether these projects can move beyond key use cases and into general-purpose production environments.
DeepSeek stated that it will continue building an open software ecosystem together with the community and will continue advancing technological innovation. Huawei's contribution to the collaboration was prominently acknowledged by the official release, but the announcement did not provide any independent response from NVIDIA, third-party organizations, or users. As the code becomes available, industry attention is expected to focus on the completeness of the documentation, interface compatibility, the frequency of ongoing maintenance, and reproducible benchmark performance. Whether independent analysts and developers can successfully migrate existing workloads to Ascend will be an important metric for evaluating the value of this release.
From an industry-trend perspective, competition in AI infrastructure is expanding from individual chip performance to full-stack capabilities encompassing compilers, operators, communication libraries, and developer tools. NVIDIA has long established relatively high migration costs through CUDA, while other accelerator platforms need to narrow the gap through coordinated software and hardware development. DeepSeek's open-source release shows that model companies and chip manufacturers are working together more deeply to define infrastructure software. For Ascend, a comprehensive open-source ecosystem may enhance the appeal of domestic computing power for large-model training and inference. For the industry as a whole, open and reusable software components can also help reduce the cost of developing multiple hardware platforms in parallel.
