DeepSeek open-sources six Huawei Ascend projects, achieving 99.8% chip peak speed
Chinese AI company DeepSeek has open-sourced six infrastructure components for Huawei's Ascend computing platform, including the TileLang programming language. The matrix kernels achieve 99.8% of the chip's theoretical peak speed. The collaboration aims to reduce reliance on Nvidia's CUDA ecosystem amid US export restrictions. DeepSeek founder Liang Wenfeng told investors Huawei could begin delivering training chips this quarter.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Common ground
- The missing license on DeepEP-Ascend is a real legal barrier that needs fixing, not just a minor oversight.
- US export controls have pushed China to build its own AI hardware ecosystem, even if it's not a free market choice.
- DeepSeek's 99.8% peak speed figure is a marketing signal, not proof of real-world performance at scale.
- Trust in either Nvidia's or Huawei's ecosystem is still lacking, and neither has fully earned it yet.
- The global south is watching this development pragmatically, not ideologically, and will adopt what works.
Points of contention
- Whether this is a genuine technical breakthrough or just a strategic hedge driven by export controls.
- Whether the missing license is a fixable documentation issue or a fundamental failure of the release process.
- Whether Huawei's ecosystem offers real freedom or just swaps one dependency for another with state control.
- Whether DeepSeek is a private company making rational survival choices or a tool of state-directed industrial policy.
- Whether the global south sees this as an escape hatch from US dominance or a choice between two flawed systems.
Blind spots
- No one has provided independent verification of sustained training throughput for DeepSeek's V4 model on Ascend hardware.
- The debate ignored how smaller developers in non-aligned countries might actually use this stack in practice.
- There was little discussion of how quickly the licensing issue on DeepEP-Ascend could be resolved and what that would mean.
- The long-term reliability and community support for TileLang compared to CUDA's decade of optimization was not explored.
- The potential for this ecosystem to evolve beyond state backing into a truly open market alternative was overlooked.
WorldAttention’s read
This debate shows that DeepSeek's open-source release for Huawei's Ascend platform is a significant strategic move, but it's not yet a proven technical breakthrough. The missing license on a core component is a real legal hurdle, and the 99.8% peak speed figure is just a marketing claim until we see independent, sustained training results. While US export controls have clearly pushed China to build its own ecosystem, it's still unclear if this will offer real freedom or just swap one dependency for another. The global south is watching pragmatically, and trust in either Nvidia's or Huawei's system remains low. Ultimately, this is a hedge born of necessity, not a clean victory for open-source alternatives.
Reporting timeline
DeepSeek open-sources six Huawei Ascend projects with matrix kernels at 99.8% chip peak speed
DeepSeek has open-sourced six projects for Huawei's Ascend AI chips, including TileLang, a programming language for writing GPU kernels. The matrix kernels in these projects achieve 99.8% of the chip's peak speed. Most operators used to train DeepSeek's V4 models are written in TileLang, which now has an Ascend version, allowing the same source code to compile for Huawei chips instead of Nvidia ones. A second repository, DeepGEMM-Ascend, performs matrix multiplication and reports 431 BF16 TFLOPS against a theoretical ceiling of 432. DeepEP-Ascend, which handles inter-chip data movement, was shipped without a license file, making it legally non-reusable. DeepSeek founder Liang Wenfeng told investors that Huawei could begin delivering training chips this quarter, a move that could further reduce reliance on Nvidia.
DeepSeek releases open-source toolkit for Huawei Ascend chips, challenging Nvidia CUDA
DeepSeek has released an open-source toolkit for Huawei's Ascend accelerators, including TileLang support optimized for the Ascend 950 chip. The tools simplify writing the compute kernels behind AI models, with the compiler handling tasks such as scheduling and synchronization. This development could make Huawei hardware more practical for developers building outside Nvidia's CUDA ecosystem. The post, citing Bloomberg, frames this as part of China's broader effort to reduce dependence on Nvidia chips and ramp up its domestic AI chip industry. The author emphasizes that while China is still dependent on Nvidia, it is taking steps to break free from this reliance.
Read sourceDeepSeek partners with Huawei to develop chip programming tools, reducing reliance on Nvidia
Chinese AI company DeepSeek announced on Wednesday a partnership with Huawei Technologies to develop programming tools for Huawei's Ascend chips, marking a significant step in China's efforts to build an alternative to Nvidia's dominant ecosystem. DeepSeek stated on its official WeChat account that it is open-sourcing infrastructure for the Huawei Ascend platform, including a high-level programming language called TileLang, along with related computing and communication libraries. The company said TileLang offers a 'simpler programming model' than Nvidia's CUDA, aiming to improve development efficiency and simplify code logic. This collaboration reflects deepening ties between Chinese tech firms as they seek to reduce dependence on Nvidia's hardware and software ecosystem amid ongoing US export restrictions on advanced semiconductors to China.
Read sourceShow 2 older updatesHide older updates
DeepSeek Open-Sources Ascend Infrastructure Components for Huawei Platform
On September 30, DeepSeek announced the open-sourcing of infrastructure components for Huawei's Ascend computing platform, including the TileLang high-level language compiler, a compute library, and a distributed communication library. These components mirror those previously released for Nvidia's platform. DeepSeek stated that Huawei provided 'unreserved strong support' during the development process, and the two companies collaborated closely on a 128-card supernode solution based on the Ascend 950, achieving deep optimization of computation and communication. The announcement emphasizes continued technological innovation and a commitment to building an open software ecosystem with the community.
DeepSeek Open-Sources Infrastructure Components for Huawei Ascend Computing Platform
DeepSeek, an AI company, has announced the open-sourcing of infrastructure components specifically designed for Huawei's Ascend computing platform. The released components include TileLang, a compilation tool; DeepGEMM; DeepEP; TileKernels; FlashMLA; and DeepSelect. These components correspond to previously open-sourced components for the Nvidia platform, indicating an effort to expand software ecosystem support for Huawei's AI hardware. The announcement, made via DeepSeek's official WeChat public account, provides functional descriptions and open-source links for each component, allowing developers to assess the usability of the Ascend platform's software ecosystem.
Read source