DeepSeek and Huawei Release Open-Source CUDA Alternative for Ascend Chips

DeepSeek and Huawei have open-sourced a suite of tools for Huawei Ascend AI accelerators, aiming to reduce dependency on Nvidia's CUDA ecosystem.
DeepSeek and Huawei Release Open-Source CUDA Alternative for Ascend Chips
The companies released an open-source suite of tools for Huawei Ascend chips. Currently, these are the primary Chinese AI accelerators that startups are migrating to. The major challenge with this hardware transition lies in software: the entire industry has been writing code for Nvidia's CUDA for over a decade, and now China faces the task of rebuilding that ecosystem for its own chips.
This is precisely what DeepSeek is addressing. Specifically, they have open-sourced Ascend versions of their proprietary libraries, which underpin their own models:
- DeepGEMM (matrix multiplication)
- DeepEP (inter-chip communication for MoE models)
- FlashMLA (attention mechanisms)
The APIs for these libraries are identical to their Nvidia counterparts, allowing code to migrate from Nvidia hardware with minimal rewriting.
The entire stack runs on TileLang, another Chinese project. DeepSeek claims that TileLang is more compact than CUDA, and all TileLang kernels used for training their models are already operational on Ascend.
While this is not a replacement for the entire CUDA stack, but rather a tool for writing kernels, the direction is clear: China is methodically eliminating its dependency on Nvidia at all levels.
Why it matters
AnalysisThis development highlights China's strategic push toward sovereign AI infrastructure, potentially challenging Nvidia's dominance and reshaping the global AI hardware landscape.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
DeepSeek open-sources part of its chat interface
DeepSeek has released the source code for a portion of its chat UI, including web search, deep thinking, and conversation history features.
DeepSeek (Chinese Breakthrough in Open AI)
A series of groundbreaking open models from the Chinese lab DeepSeek (V3, R1). Demonstrates top-tier coding and reasoning capabilities at a cost 10–20 times lower than Western commercial counterparts.
Nvidia's Monopoly and the CUDA Platform
An analysis of Nvidia's technological and economic dominance in the AI market. The CUDA (Compute Unified Device Architecture) platform, created in 2006, transformed ordinary gaming GPUs into the planet's primary computational tool, making it difficult for competitors like AMD and Intel to break this monopoly.
DeepSeek app capabilities now available through API
DeepSeek exposes search, file and link handling, safety functions and voice capabilities through its API, plus an open-source demo interface for building chatbots.