DeepSeek Open Sources Full Ascend Stack, Intensifying AI Hardware Ecosystem Competition

In a move that could reshape the competitive landscape, AI research company DeepSeek open-sourced its complete suite of infrastructure components for Huawei's Ascend AI computing platform on September 30th. This release provides Ascend ecosystem developers with a software toolkit that mirrors the capabilities available for NVIDIA platforms, signaling a new phase of competition in the foundational AI software layer.

A Complete Development Suite: From Compilation to Communication

The release centers on three integrated components: the TileLang high-level language compilation tool, a set of core computing libraries, and distributed communication libraries optimized for large-scale training. Together, they form a full-stack solution for developing efficient AI applications on the Ascend platform.

TileLang: Elevating Programming Abstraction for Ascend

The TileLang compiler is a cornerstone of the release for Ascend. It provides a high-level abstraction over the Ascend C low-level instructions, allowing developers to write code in a more intuitive manner. Its sophisticated compilation techniques ensure the generated programs fully leverage the hardware's capabilities, eliminating the performance penalty often associated with high-level programming.

Five Core Components for Key AI Workloads

Beyond the compiler, the open-source library includes five core modules optimized for distinct computational patterns:

  • DeepGEMM: Accelerates general matrix multiplication, one of the most fundamental and time-consuming operations in deep learning models.
  • DeepEP: Engineered for large-scale distributed training, it enables efficient cross-device and cross-node communication to minimize training bottlenecks.
  • TileKernels: Offers a collection of essential vector computation and memory access operators for data processing, serving as building blocks for more complex operations.
  • FlashMLA: Optimizes the sparse attention mechanism—a critical component for modern large language models—significantly boosting efficiency in long-context processing tasks.
  • DeepSelect: Implements high-performance data selection operations, crucial for data preprocessing and specific model inference scenarios.

Performance: Pushing Hardware to Its Limits

According to DeepSeek's benchmarks, these components demonstrate exceptional performance in key tests. Their computational efficiency and communication bandwidth utilization are reported to approach the theoretical limits of the Ascend hardware. This allows developers to achieve near-peak hardware performance without engaging in complex low-level optimizations, substantially lowering the barrier to building high-performance AI applications on heterogeneous computing platforms.

This open-source move clarifies DeepSeek's technical strategy. By providing performance-competitive, feature-complete software stacks for different mainstream AI chip platforms, it is positioning itself to set a new industry standard in foundational AI software. For developers and enterprises, this increases flexibility and control in their hardware selection process.