DeepSeek Huawei Ascend Partnership
DeepSeek and Huawei Open-Source Software Tools to Challenge Nvidia Ecosystem on Ascend Chips
Anonymous
DeepSeek and Huawei are jointly developing and open-sourcing software for Huawei’s Ascend AI chips, including TileLang programming tools, compute libraries and software for communication between chips. The effort also includes work on a supernode combining 128 Ascend 950 chips, though the companies have not released independent performance results for the full system. DeepSeek says TileLang aims to make it easier to write high-performance AI programs, offering developers an alternative to Nvidia’s CUDA software ecosystem. The partnership addresses a key challenge for China’s domestic AI chips: building software and developer tools that help them compete with established platforms, amid restrictions on access to advanced U.S. processors. Separately, Huawei announced the release of training code for its openPangu-2.0 model, another move to expand the Ascend ecosystem.
Key Points
1
DeepSeek said TileLang was originally developed by Peking University researchers and that it had used the language for about a year; it had first tested TileLang on older Nvidia chips and now uses it as its main tool for artificial-general-intelligence work.
2
DeepSeek described TileLang as a universal, high-level language that should be easy to program while still reaching the hardware’s full performance potential.
3
Huawei’s openPangu Training Framework is designed to support large language, vision-language and multimodal models, with capabilities spanning pre-training, supervised fine-tuning and ongoing model optimization.
4
The DeepSeek announcement described the released Ascend components as corresponding one-to-one with open-source components previously released for Nvidia’s platform.
Sentiment
Mostly PositiveNEGATIVENEUTRALPOSITIVE
Publishers
34
Articles
140
Reach
174
A
AnonymousAuthor
