DeepSeek V5 to be trained on Huawei Ascend 950 AI accelerators

DeepSeek is planning to leverage Huawei Ascend 950 accelerators for the pre-training phase of its upcoming large language model generation, known as DeepSeek V5. This move highlights a major shift toward domestic hardware for advanced AI development.

• DeepSeek V5 will be the company's first major model pre-trained on Huawei hardware;
• The Ascend 950 lineup consists of two models: the 950DT and 950PR;
• The 950DT chip features 144 GB of HBM memory with a 4 TB/s bandwidth;
• MXFP4 performance is rated four times higher than the previous BF16 generation.

Architecture and Specifications of Ascend 950 The Huawei Ascend 950 series is built upon the third-generation Da Vinci architecture, specifically tailored to handle massive artificial intelligence models and high-performance computing workloads. The Ascend 950DT accelerator is optimized primarily for training and inference decoding tasks. It is equipped with 144 GB of Huawei's proprietary HBM memory, delivering a memory bandwidth of 4 TB/s and an interconnect bandwidth of 2 TB/s.

Advanced Performance and Format Support Alongside high bandwidth capabilities, the new accelerators support modern FP8, MXFP8, and MXFP4 data formats. According to published metrics, performance in the MXFP4 format is four times higher than the previous generation running in BF16. The second accelerator, the Ascend 950PR, is designed mainly for the data pre-fill stage during inference. It comes with 128 GB of embedded memory offering a 1.6 TB/s bandwidth, while delivering a computational performance of up to 1,784 Tflops in the FP4 format.

Implications for the AI Hardware Market Previously, Chinese hardware was predominantly utilized by DeepSeek for inference tasks—running already trained neural networks. Shifting the pre-training phase of a flagship model like V5 to the Ascend platform marks a significant milestone for domestic computing solutions. This development signals the growing maturity of local semiconductor ecosystems in providing robust infrastructure for cutting-edge AI model training and deployment.