Open-Source Ascend Tools
Chinese artificial intelligence company DeepSeek has released open-source programming tools developed in partnership with Huawei for Ascend AI chips, as part of broader efforts to reduce industry reliance on Nvidia's hardware and software ecosystem. The release includes open-source libraries for AI computation and chip-to-chip communication, alongside Ascend support for the high-level programming language TileLang.
The utilities, which received full development support from Huawei, aim to simplify programming while enabling engineers to fully leverage the underlying hardware performance. Both companies collaborated to optimize computation and communication on a supernode architecture built around 128 Ascend 950 processors, addressing the dual challenge of running large workloads across multiple accelerators efficiently and maintaining fast data transfer speeds.
Libraries and Communication
The newly released software includes DeepGEMM-Ascend, which manages matrix multiplication and calculations utilized within DeepSeek models, supporting BF16, FP8, and FP4 operations via familiar application programming interfaces. Additionally, DeepEP-Ascend handles communication for both training and inference workloads, managing data routing for mixture-of-experts models. Both packages were tested extensively on Ascend 950 hardware.
TileLang serves as the higher-level framework for authoring optimized kernels, offering a streamlined programming model compared to Nvidia's CUDA. While previous iterations supported various architectures, the update introduces native integration for the Ascend 950, delivering code generation, automatic scheduling, and synchronization capabilities.
Ecosystem Impact
These utilities build directly upon Huawei's CANN software stack, which supplies the underlying infrastructure for executing artificial intelligence tasks on Ascend silicon. Nvidia's CUDA platform has traditionally maintained a dominant market position due to its mature development environment, but cross-platform frameworks like TileLang ensure developers retain flexibility across different hardware architectures.
This rollout follows Huawei's recent introduction of next-generation processors and supernode systems, with commercial deployment for model training anticipated in 2027. The two organizations previously cooperated on previewing the DeepSeek V4 model in April, utilizing Ascend chips for portions of the training process.
