Full Breakdown
DeepSeek Open-Sources Full Toolchain for Huawei Ascend AI Chips
By Drooid · · How we work
Core Event: Open-Source Release on September 30
On September 30, DeepSeek announced the open-source release of its complete programming infrastructure for Huawei’s Ascend AI computing platform. The suite includes the TileLang high-level language, its compiler, compute libraries, and distributed-communication libraries, extending the same toolchain that DeepSeek previously made available for Nvidia GPUs to Huawei’s Ascend hardware.
Background & Context
Huawei’s Ascend line of AI processors is positioned as a domestic alternative to Nvidia’s GPU ecosystem. Earlier in 2026, Huawei unveiled new-generation Ascend processors and supernode systems at its Huawei Connect event, signaling a push toward a self-contained AI software stack. DeepSeek, a Chinese AI startup, has partnered with Huawei to provide the software layer that enables developers to train and run large-scale models on Ascend chips, directly addressing the need for a high-performance, open-source programming model.
Key Components Released
- TileLang – a high-level, open-source language that abstracts low-level Ascend C instructions while preserving full hardware performance.
- TileLang compiler support – translates TileLang code to optimized Ascend binaries.
- Compute libraries – including DeepGEMM for general matrix multiplication acceleration.
- Communication libraries – featuring DeepEP for large-scale cross-device data exchange.
- TileKernels – a set of conventional vector-compute and memory-access operators.
- FlashMLA – a sparse-attention operator that improves long-context processing efficiency.
- DeepSelect – tools for efficient data selection during training.
These components are described as providing “one-to-one parity” with the equivalents already released for Nvidia GPUs, meaning each TileLang operator used in DeepSeek’s V4-series models now has a high-performance Ascend implementation.
Joint Supernode Development
DeepSeek and Huawei are co-developing a supernode architecture that aggregates 128 Ascend 950 chips. The joint effort focuses on deep optimization of both compute and communication pathways, aiming to maximize throughput and approach the hardware’s theoretical performance limits in key benchmark tests.
Data & Statistics
- 128-card supernode built on Ascend 950 processors.
- TileLang operators on Ascend achieve “high-performance implementations” matching those on Nvidia GPUs.
- Early test cases show compute and communication performance “already approaching hardware limits.”
Why It Matters
The release expands the open-source ecosystem surrounding Huawei’s Ascend processors, providing developers with a programming model that rivals Nvidia’s CUDA without requiring proprietary tools. By delivering a full stack—from language to low-level kernels—DeepSeek aims to lower the barrier for AI research and deployment on domestic hardware, potentially reshaping the balance of AI accelerator adoption in China and beyond.