DeepSeek and Huawei release open-source Ascend AI programming tools to reduce reliance on Nvidia ecosystem
DeepSeek and Huawei have released open-source programming tools for Ascend 950 AI chips, including compute and communication libraries, aimed at making Huawei hardware easier to program and optimize.
Chinese AI firm DeepSeek and semiconductor maker Huawei have jointly launched open-source programming resources for DeepSeek's Ascend AI chips. This move aims to lessen dependence on Nvidia's software and hardware infrastructure, as reported by Reuters on September 30. The tools encompass libraries for AI computation and inter-chip communication, and support the TileLang high-level programming language.
Partnered on these tools, Huawei provided comprehensive support during development. They focused on optimizing calculations on individual chips and accelerating data transfer between them for efficient large-scale AI workloads. Key among these tools is DeepGEMM-Ascend, adept at matrix multiplication and handling operations like BF16, FP8, and FP4, mirroring DeepSeek's existing DeepGEMM library's programming interfaces.
Additionally, DeepEP-Ascend manages the communication required for model training and inference, including routing data to experts in mixture-of-experts models and aggregating their outputs. Both tools were tested on Ascend 950 hardware. TileLang simplifies programming by offering a more straightforward model compared to Nvidia's CUDA, aimed at enhancing development efficiency and coding simplicity.
It already accommodates Nvidia and other hardware, with earlier Ascend adapters in place. The update adds native support for Ascend 950, encompassing code generation, automatic scheduling, and synchronization. These resources leverage Huawei's existing CANN software platform, offering a robust infrastructure for running AI workloads on Ascend.
While DeepSeek's release offers developers more tools to optimize Ascend workloads, TileLang's versatility across platforms ensures its relevance for Nvidia GPUs as well. This announcement follows closely after Huawei's unveiling of a new generation of AI processors and supernode systems, with expectations of widespread use of these systems for model training by 2027.
Prior to this, DeepSeek and Huawei collaborated on DeepSeek's V4 model, which ran on Ascend chips and was previewed in April.
Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
