DeepSeek open-sources Huawei Ascend infrastructure components matching its NVIDIA stack, IT Home reports
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
DeepSeek open-sourced infrastructure components for Huawei Ascend, including TileLang compiler, compute libraries, and distributed communication libraries, IT Home reported. The components correspond one-to-one with its NVIDIA platform releases. TileLang Ascend version wraps Ascend C instructions without hardware performance loss.
China context
- Original name
- 深度求索
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can evaluate TileLang Ascend for high-performance operator development on Huawei hardware; verify licensing and compatibility with existing Ascend toolchains.
- For investors
- This release may strengthen Huawei Ascend's software ecosystem, potentially affecting demand for NVIDIA alternatives in China and among export-restricted customers.
Translated from Chinese. Quotes and facts link to the original sources.
DeepSeek announced via its official WeChat account the open-sourcing of infrastructure components for Huawei Ascend compute platform, covering TileLang high-level language compiler, compute libraries, and distributed communication libraries. All components correspond one-to-one with previously open-sourced components for NVIDIA platform. TileLang is described as simpler than CUDA, improving development efficiency and simplifying code logic, while fully utilizing chip characteristics to reach hardware performance limits. The TileLang route was first validated on NVIDIA mature platform and currently carries most operator implementations in DeepSeek V4 series model training. The open-sourced Ascend version encapsulates Ascend C underlying instructions, provides high-level language programming, and does not lose hardware performance. Every TileLang operator used in DeepSeek training has a corresponding high-performance implementation on Ascend. The release also includes core compute and communication components: DeepGEMM for general matrix operations, DeepEP for large-scale cross-device communication, TileKernels for conventional vector computation and memory access operators, FlashMLA for sparse attention operators to improve long-context processing efficiency, and DeepSelect for efficient data filtering. In multiple key test cases, compute and communication performance of these components is close to hardware limits.
TileLang provides a high-level programming model that abstracts Ascend C instructions, aiming to simplify development while preserving hardware performance. The release includes DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, covering matrix operations, communication, vector compute, sparse attention, and data selection. DeepSeek states that every TileLang operator used in its training has a corresponding high-performance Ascend implementation, and that performance in key tests is near hardware limits. These claims are company-reported; independent benchmarks are not provided in the evidence.
Developers targeting Huawei Ascend hardware gain access to DeepSeek's training-validated operator stack, potentially reducing the effort to port or build high-performance AI workloads on Ascend. This could lower the barrier for organizations using Ascend for large-model training and inference, and may shift some developer mindshare from NVIDIA CUDA to Ascend for certain workloads.
For enterprises and developers invested in Huawei Ascend infrastructure, this release provides a set of open-source, training-proven components that may reduce development cost and time for high-performance AI applications. It also offers an alternative to NVIDIA CUDA-based tooling for organizations subject to export controls or preferring domestic Chinese hardware.
A verifiable next signal is whether DeepSeek or third parties publish independent benchmarks comparing Ascend and NVIDIA implementations of these components. Another signal is adoption of TileLang Ascend in public projects or model training runs outside DeepSeek. The evidence does not state availability outside China or licensing terms beyond open source.