DeepSeek and Huawei have introduced open-source programming tools designed to optimize AI workloads on Ascend hardware. The collaboration aims to simplify development and reduce dependence on the Nvidia ecosystem by providing libraries that manage computation and data communication across accelerator clusters.
The new release includes DeepGEMM-Ascend, which supports BF16, FP8, and FP4 operations for matrix multiplication, and DeepEP-Ascend, a library for training and inference communication. Both tools were specifically developed and tested on systems utilizing 128 Ascend 950 chips.
The TileLang programming layer received an update on September 30 to add native support for Ascend 950 processors. This includes features for code generation, automatic scheduling, and synchronization, offering an alternative programming model to Nvidia CUDA.
These tools integrate with Huawei's existing CANN software platform. The companies previously collaborated on the DeepSeek V4 model, which was released in April, and Huawei confirmed that its Ascend 950 supernodes fully support these models for training purposes.
Huawei recently unveiled its next-generation AI processors and supernode systems. The company anticipates that its AI infrastructure will see widespread adoption for model training by 2027.
