- Huawei introduces Atlas 350 with significant FP4 compute performance claims
- New accelerator card focuses on inference workloads and multimodal AI processing
- Huawei Atlas 350 delivers higher memory capacity and improved bandwidth efficiency
Huawei has officially launched the Atlas 350 accelerator card, featuring its new Ascend 950PR processor, at the Huawei China Partner Conference 2026 in Shenzhen.
The company claims this NPU delivers 1.56 PFLOPS of FP4 compute performance, which is reportedly 2.87 times higher than Nvidia’s H20.
While exact verification is difficult because Hopper-era GPUs do not support FP4 natively, the Atlas 350 is the first Chinese accelerator optimized for this low-precision format, allowing larger AI models to operate on the same hardware with reduced memory requirements.
Article continues below
Technical upgrades and memory performance
The Ascend 950PR chip introduces improvements over the prior Ascend 910 series, including enhanced microarchitecture, faster memory access, and flexible programming modes.
Huawei equips the Atlas 350 with 112GB of proprietary HBM, known as HiBL 1.0, delivering up to 1.4TB/s bandwidth in current reports, with a 128-byte memory access granularity.
This configuration enables efficient multimodal generation and inference tasks, and reportedly quadruples memory access efficiency for small operators compared with the previous generation.
Its interconnect bandwidth also reaches 2TB/s using the LingQu protocol, 2.5 times higher than the Ascend 910 series.
Huawei markets the Atlas 350 for recommendation inference, LLM processing, and multimodal AI workloads.
Seven key partners — including Kunlun, Huakun Zhenyu, Shenzhou Kuntai, and Yangtze Computing — have developed complete system products leveraging the Atlas 350.
These brands have created customized high-performance inference solutions for enterprise customers.
The accelerator is designed to integrate with AI ecosystems, enabling partners to optimize performance for specific workloads while maintaining compatibility with Huawei’s AI software stack.
The Atlas 350 reflects China’s efforts to establish self-reliance in AI compute hardware under U.S. export restrictions.
While Huawei cannot access TSMC’s CoWoS technology, the company has implemented alternative advanced packaging solutions for HBM and memory stacking.
Huawei has not announced precise availability dates — a common practice with AI accelerators — but it launched the Ascend 950PR in Q1 2026 as promised.
The Atlas 350 is reportedly priced at around 111,000 Yuan, or roughly $16,000, comparable with Nvidia H20, which can range from $15,000 to $25,000.
Via Tom’s Hardware
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
https://cdn.mos.cms.futurecdn.net/MBbjHBboVJMsz3Hb8xpA44-1920-80.png
Source link
