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 neural processing unit (NPU) delivers 1.56 PFLOPS of FP4 compute performance, reportedly 2.87 times higher than Nvidia’s H20. While exact verification is challenging due to Hopper-era GPUs lacking native FP4 support, the Atlas 350 is noted as the first Chinese accelerator optimized for this low-precision format, which allows larger AI models to function on the same hardware with reduced memory requirements.
The Ascend 950PR chip brings significant improvements over the previous Ascend 910 series. Key enhancements include a revamped microarchitecture, accelerated memory access, and flexible programming modes. Huawei has equipped the Atlas 350 with 112GB of proprietary high-bandwidth memory (HBM), known as HiBL 1.0, which reportedly enables bandwidth of up to 1.4TB/s and features a 128-byte memory access granularity. This configuration is designed to facilitate efficient multimodal generation and inference tasks, and it purportedly quadruples memory access efficiency for small operators compared to its predecessor.
Interconnect bandwidth also reaches 2TB/s using the LingQu protocol, representing a 2.5-fold increase over the Ascend 910 series. The Atlas 350 is specifically marketed for recommendation inference, large language model (LLM) processing, and multimodal AI workloads. Seven key partners, including Kunlun, Huakun Zhenyu, Shenzhou Kuntai, and Yangtze Computing, have developed complete system solutions leveraging the Atlas 350, creating customized high-performance inference products for enterprise customers.
The accelerator card is designed to integrate seamlessly with AI ecosystems, empowering partners to optimize performance for specific workloads while maintaining compatibility with Huawei’s AI software stack. The Atlas 350 also reflects China’s broader strategy to enhance self-reliance in AI computing hardware amidst ongoing U.S. export restrictions. Despite Huawei’s inability to access TSMC’s CoWoS technology, the company has implemented alternative advanced packaging solutions for HBM and memory stacking to bolster its offerings.
While Huawei has not disclosed precise availability dates—a common practice for AI accelerators—it successfully launched the Ascend 950PR in Q1 2026 as scheduled. The Atlas 350 is expected to be priced at around 111,000 Yuan, or approximately $16,000, which positions it competitively with Nvidia’s H20, priced between $15,000 and $25,000. The introduction of the Atlas 350 could represent a pivotal step in Huawei’s quest to carve out a more significant role in the global AI hardware landscape, particularly at a time when demand for advanced computing solutions continues to surge.
See also
Google’s Labs Reveals Hidden AI Tools Like NotebookLM with Impressive Features
OpenAI Abandons Controversial Adult Chatbot Amid Legal Concerns and Investor Pushback
Hugging Face Launches Smolagents for Effortless Python AI Agent Development in 15 Minutes
AGI’s Last Invention Claim Challenges: Human Expertise Remains Crucial, Says Expert
Anthropic Wins Preliminary Injunction Against Pentagon, Legal Risks Remain Uncertain
