DeepSeek Opens Source Full-Stack Toolchain for Ascend, Filling Gaps in Domestic Compute Ecosystem

Overview

Recently, the domestic large model developer DeepSeek announced the open-sourcing of a complete Ascend-based foundational toolchain. This move fills part of the software development gap in the Huawei Ascend ecosystem, aiming to lower the barrier for developers building model training and inference on domestic computing power. The toolchain mirrors DeepSeek's previously open-sourced NVIDIA CUDA version components, marking a new stage in the development of China's domestic AI infrastructure ecosystem.

Core Components and Technical Breakthroughs

The open-sourced toolchain covers multiple key links from low-level operator writing to high-level communication optimization, mainly including the following components:

• TileLang Ascend Edition: As the core framework, it provides a high-level wrapper around low-level Ascend C instructions, allowing developers to write operators using more user-friendly high-level languages without directly grappling with complex low-level assembly instructions.
• DeepGEMM: Focuses on accelerating matrix multiplication by optimizing the computational efficiency of general matrix multiplication.
• DeepEP: Deeply optimizes cross-card communication for large-scale clusters, addressing data synchronization bottlenecks in multi-card collaborative training.
• TileKernels: Provides high-performance implementations for standard vector computations and memory-access operators.
• FlashMLA: Introduces a sparse attention mechanism, significantly improving runtime speed in long-context scenarios.
• DeepSelect: Implements an efficient data filtering mechanism, optimizing the data pre-processing pipeline.

"Every TileLang operator used in DeepSeek's training has a corresponding high-performance implementation on the Ascend platform."

Official test data shows that several key performance metrics are approaching hardware limits, demonstrating extremely high engineering implementation value.

Industry Impact and Ecosystem Significance

For a long time, the challenge faced by domestic AI chips has not been limited to computing power itself, but more importantly, the scarcity of a software ecosystem. NVIDIA's long-standing market dominance is largely attributable to the moat built by its CUDA ecosystem. DeepSeek's open-source initiative this time is viewed by the industry as a powerful patch for the shortcomings of China's domestic computing power ecosystem.

By open-sourcing core infrastructure, DeepSeek has reduced the costs for other research institutions and enterprises to adapt to the Ascend platform. This move not only enhances the software usability of domestic hardware but also lays an important foundation for building an autonomous and controllable AI computing power ecosystem. Analysts point out that this strategy of "doing what Huawei should have done first" is essentially accelerating the leap of domestic computing power from "usable" to "easy to use."