- Huawei introduces Atlas 350 with vital FP4 compute efficiency claims
- New accelerator card focuses on inference workloads and multimodal AI processing
- Huawei Atlas 350 delivers greater reminiscence capability and improved bandwidth effectivity
Huawei has formally launched the Atlas 350 accelerator card, that includes its new Ascend 950PR processor, on the Huawei China Accomplice Convention 2026 in Shenzhen.
The corporate claims this NPU delivers 1.56 PFLOPS of FP4 compute efficiency, which is reportedly 2.87 instances greater than Nvidia’s H20.
Whereas precise verification is troublesome as a result of Hopper-era GPUs don’t assist FP4 natively, the Atlas 350 is the primary Chinese language accelerator optimized for this low-precision format, permitting bigger AI fashions to function on the identical {hardware} with lowered reminiscence necessities.
Article continues under
It’s possible you’ll like
Technical upgrades and reminiscence efficiency
The Ascend 950PR chip introduces enhancements over the prior Ascend 910 collection, together with enhanced microarchitecture, sooner reminiscence entry, and versatile programming modes.
Huawei equips the Atlas 350 with 112GB of proprietary HBM, often called HiBL 1.0, delivering as much as 1.4TB/s bandwidth in present stories, with a 128-byte reminiscence entry granularity.
This configuration permits environment friendly multimodal technology and inference duties, and reportedly quadruples reminiscence entry effectivity for small operators in contrast with the earlier technology.
Its interconnect bandwidth additionally reaches 2TB/s utilizing the LingQu protocol, 2.5 instances greater than the Ascend 910 collection.
Huawei markets the Atlas 350 for suggestion inference, LLM processing, and multimodal AI workloads.
Seven key companions — together with Kunlun, Huakun Zhenyu, Shenzhou Kuntai, and Yangtze Computing — have developed full system merchandise leveraging the Atlas 350.
These manufacturers have created custom-made high-performance inference options for enterprise clients.
What to learn subsequent
The accelerator is designed to combine with AI ecosystems, enabling companions to optimize efficiency for particular workloads whereas sustaining compatibility with Huawei’s AI software program stack.
The Atlas 350 displays China’s efforts to determine self-reliance in AI compute {hardware} underneath U.S. export restrictions.
Whereas Huawei can’t entry TSMC’s CoWoS expertise, the corporate has applied different superior packaging options for HBM and reminiscence stacking.
Huawei has not introduced exact availability dates — a typical follow with AI accelerators — however it launched the Ascend 950PR in Q1 2026 as promised.
The Atlas 350 is reportedly priced at round 111,000 Yuan, or roughly $16,000, comparable with Nvidia H20, which may vary from $15,000 to $25,000.
Through Tom’s {Hardware}
Comply with TechRadar on Google Information and add us as a most popular supply to get our skilled information, opinions, and opinion in your feeds. Make sure that to click on the Comply with button!
And naturally you can even observe TechRadar on TikTok for information, opinions, unboxings in video type, and get common updates from us on WhatsApp too.
