DeepSeek Adapted to Ascend 950: Inference Deployed, Training Still in Progress

Overview

On September 30, DeepSeek did not release a new pre-trained model, but a series of major progress reports regarding its core model adaptation to Huawei's Ascend 950 infrastructure were officially released. This update covers seven key projects ranging from low-level matrix operations to upper-layer model architectures, marking a phased breakthrough in DeepSeek's deployment on domestic computing power clusters.

The focus of this release lies in optimizing the compiled runtime support for core components such as matrix multiplication, sparse attention mechanisms, and MoE (Mixture of Experts) communication. This means that DeepSeek's flagship models can now achieve efficient inference services on Ascend 950 hardware, laying a solid foundation for subsequent large-scale commercial implementation.

Inference Deployment: Performance Approaching Theoretical Limits

According to technical details disclosed in the DeepSeek official repository, the current adaptation work has achieved substantive results first in the inference stage. After deep operator optimization, DeepSeek's GEMM (General Matrix Multiply) performance on Ascend 950 has climbed to 99.8% of the theoretical peak computing power. This data indicates that the computational potential of the underlying hardware has been nearly fully released, significantly reducing performance losses caused by software stack overhead.

"This GEMM performance breakthrough of 99.8% means that our optimization at the operator level has become mature, capable of providing stable and efficient computing power support for upper-layer applications." —— DeepSeek Technical Team

In addition to the optimization of dense models, this update also focuses on solving the compatibility issue of sparse attention. For large models like DeepSeek-V3 that adopt long context windows, the sparse attention mechanism can significantly reduce video memory usage and improve processing speed. Currently, this feature has been verified on the Ascend 950 cluster, and inference latency has improved significantly compared to before.

Training Breakthroughs: MoE Communication and Compiler Optimization

Although progress on the inference side has been smooth, the adaptation of the training process is still proceeding in full swing. DeepSeek's V3 and R1 series models adopt a complex MoE architecture, which places extremely high demands on inter-node communication during distributed training. Among the updates released this time, the optimization of the MoE communication protocol is one of the key breakthroughs.

Specifically, the team reconstructed the compiler runtime to better adapt to the Ascend CANN software stack. This includes optimizing the All-to-All communication pattern to reduce network congestion during the expert routing process and improving parallel efficiency under multi-card interconnection. Although training performance has not yet reached the level of the inference side, the core communication bottleneck has initially been alleviated, and test versions of training capabilities are expected to be further opened up in the near future.

Future Prospects

This batch of updates not only reflects DeepSeek's high attention to the domestic computing power ecosystem, but also demonstrates the deep collaboration between the two parties in heterogeneous computing optimization. With the maturity of inference capabilities, it is expected that more inference services based on DeepSeek models will go online on the Ascend 950 cluster, providing users with lower latency and higher concurrency processing capabilities. On the training end, with the deepening of communication optimization, an end-to-end autonomous and controllable training closed loop is expected to be realized in the future, further enriching the technical foundation for domestic AI development.