Home / Technology / Huawei Unleashes AI Powerhouse: Atlas 350
Huawei Unleashes AI Powerhouse: Atlas 350
29 Mar
Summary
- Atlas 350 boasts 1.56 PFLOPS FP4 compute, outperforming Nvidia H20.
- It offers 112GB HBM and 2TB/s interconnect for enhanced AI tasks.
- The accelerator aims to boost China's AI hardware self-reliance.
Huawei has officially unveiled the Atlas 350 accelerator card, integrating its new Ascend 950PR processor. The company states this NPU delivers 1.56 PFLOPS of FP4 compute performance, a significant advancement. This focus on FP4 allows for larger AI models to operate with reduced memory demands.
The Atlas 350 is equipped with 112GB of proprietary HBM, HiBL 1.0, offering substantial bandwidth and improved memory access efficiency. Its interconnect bandwidth reaches 2TB/s, facilitating multimodal generation and inference tasks. Huawei is marketing this accelerator for recommendation inference, large language models, and multimodal AI applications.
This launch is seen as a key part of China's strategy to achieve self-reliance in AI compute hardware, particularly under current export restrictions. Several key partners have already developed system products utilizing the Atlas 350, offering customized inference solutions for enterprise customers.
The Atlas 350 is priced comparably to Nvidia's H20, with reports suggesting around 111,000 Yuan (approximately $16,000). Huawei promised the Ascend 950PR launch in Q1 2026, and it has met this timeline.
