DeepSeek joins hands with Huawei to build AI chip software, cutting Nvidia reliance

This partnership is part of a broader push by Chinese technology companies to establish domestic alternatives to foreign semiconductor hardware and software. DeepSeek's release of open-source programming infrastructure for the Ascend platform, including compute and communication libraries, addresses the need for a comprehensive software stack to support Huawei's AI chips.

A key element of the joint effort is the focus on software efficiency, highlighted by TileLang. DeepSeek describes TileLang as an open-source, high-level programming language intended for artificial intelligence chips, designed to simplify programming and enable full utilization of hardware performance. This suggests an attempt to offer a more accessible development environment compared to existing proprietary solutions, such as Nvidia's CUDA.

Huawei provided full technical backing throughout this development, including joint work on a supernode configuration built with 128 Ascend 950 chips. This collaboration on refining computational performance and communication links within clustered systems indicates an ambition to optimize the broader system architecture necessary for large-scale AI model training, beyond just individual chip capabilities.