DeepSeek has made the source code for six software components available for Huawei Ascend AI chips. These components replicate essential tools for Nvidia products, establishing an independent software stack for Chinese accelerators, according to SCMP.
DeepSeek Integrates Its Stack with Huawei
The components introduced by DeepSeek for the Huawei Ascend platform include the TileLang programming language, as well as the DeepGEMM and DeepEP libraries, along with TileKernels, FlashMLA, and DeepSelect. These tools are designed to align with technologies that DeepSeek previously developed for Nvidia chips.
A standout offering is the TileLang version for Ascend, which is designed for creating high-performance computing kernels—small programs executing critical operations within AI models. Bloomberg referred to it as "China's answer to Nvidia CUDA."
TileLang now officially supports Huawei Ascend 950 accelerators, featuring capabilities such as native code generation, automatic scheduling, and synchronization.
"As an open-source project, the TileLang Ascend version aims to serve as a model for building a widely accessible software ecosystem for a larger array of AI chips," stated DeepSeek.
From Matrix Computations to Data Exchange
The remaining components address crucial tasks involved in training and operating AI models.
DeepGEMM Ascend handles matrix computations and is compatible with the original DeepGEMM API. It supports BF16, FP8, and FP4 formats and is optimized for the Ascend architecture.
According to the project, individual kernels are designed to operate close to the limits of hardware performance.
DeepEP Ascend enhances data exchange among numerous chips, a vital feature for models utilizing the Mixture of Experts architecture. Other components focus on speeding up memory operations, processing long contexts, implementing sparse attention, and filtering data.
DeepSeek and Huawei Combine Chips and Software
As part of their collaboration, the companies have also developed a solution for a "supernode," which consists of 128 Huawei Ascend 950 chips and optimizes both computation and data exchange among the accelerators.
Huawei has supported DeepSeek in building the software infrastructure, and both companies plan to continue jointly developing tools for Ascend.
It is worth noting that in September, Huawei announced the upcoming release of the Ascend 960DT chips in the first quarter of 2027, ahead of the initially scheduled timeline.
Follow ForkLog on social media:
Telegram (main channel) Facebook X