AI and Open Source DeepSeek and Huawei Make Programming Tools for Ascend Chips Available

By Sebastian Gerstl | Translated by AI 2 min Reading Time

Related Vendors

DeepSeek and Huawei have released key programming tools for Ascend AI chips under an open-source license. This is intended to simplify development on Huawei's hardware and reduce dependence on NVIDIA's CUDA ecosystem.

Open-source AI tools: DeepSeek and Huawei are making programming tools for Ascend processors available, with the aim of simplifying development, chip communication, and AI computations on Huawei's hardware—and thereby intensifying competition with Nvidia.(Image: Dall-E / AI-generated)
Open-source AI tools: DeepSeek and Huawei are making programming tools for Ascend processors available, with the aim of simplifying development, chip communication, and AI computations on Huawei's hardware—and thereby intensifying competition with Nvidia.
(Image: Dall-E / AI-generated)

Chinese AI developer DeepSeek has collaborated with Huawei to develop new programming tools for Huawei’s Ascend AI chips and released them as open source. These include libraries for computational operations and inter-chip communication, as well as expanded support for the high-level language TileLang. The collaboration is part of efforts by Chinese technology companies to build alternatives to Nvidia’s previously dominant hardware and software ecosystem.

According to DeepSeek, Huawei provided extensive support for the development of the new infrastructure. Both companies also worked on optimizing a so-called supernode system with 128 Ascend-950 chips. The goal was both to efficiently perform computational operations on the individual processors and to quickly transfer large amounts of data between the chips.

Among the released components is DeepGEMM-Ascend. Among other things, the library handles matrix multiplications—such as those required for training and running large AI models—and supports various numeric formats, including BF16, FP8, and FP4. DeepEP-Ascend, on the other hand, is designed for communication between accelerators, such as when distributing and merging data in so-called mixture-of-experts models. Both libraries were developed and tested on Ascend 950 hardware.

TileLang Is Designed to Simplify Programming

TileLang, an open-source high-level language for developing optimized computing cores for AI chips, also plays an important role. According to DeepSeek, it is designed to give developers easier access to the hardware while maximizing its performance. The latest update adds native support for the Ascend 950, including features such as code generation, automatic scheduling, and synchronization.

TileLang is not exclusively designed for Huawei chips. The language also supports hardware from other vendors, including Nvidia GPUs. DeepSeek nevertheless presents the approach as a building block for a more independent software ecosystem. The new tools build on Huawei’s existing CANN platform, which provides basic software functions for running AI applications on Ascend processors.

The release comes about two weeks after Huawei unveiled a new generation of AI processors and supernode systems. The company expects its systems to be used more extensively for training AI models starting in 2027. DeepSeek and Huawei had previously collaborated on DeepSeek’s V4 model family. With the developer tools now released, the Ascend platform is set to become more widely usable on the software side as well, thereby creating an alternative to NVIDIA’s established CUDA ecosystem.

Subscribe to the newsletter now

Don't Miss out on Our Best Content

By clicking on „Subscribe to Newsletter“ I agree to the processing and use of my data according to the consent form (please expand for details) and accept the Terms of Use. For more information, please see our Privacy Policy. The consent declaration relates, among other things, to the sending of editorial newsletters by email and to data matching for marketing purposes with selected advertising partners (e.g., LinkedIn, Google, Meta)

Unfold for details of your consent