Huawei has introduced the Atlas 350 accelerator card at the Huawei China 2026 Partner Conference held in Shenzhen. It will be direct competition to NVIDIA’s H20, exclusive for China after the Trump government’s export approval.
Blocking the export of advanced chips to China has been the general rule of the US legislature under the always recurring umbrella of “national security.” For this reason, China has set the goal of achieving complete self-sufficiency in the field of artificial intelligence infrastructure. Huawei’s Atlas 350 will be a new advance in that strategy.
Atlas 350, powerful and national
This new NPU is based on a self-made Ascend 950PR chip, which represents a significant improvement over the previous generation Ascend 910 chip. This is a high-performance processor designed for the preload (inference) phase in AI implementation. Offers a performance of 1,56 PFLOPS en FP4which according to Huawei is 2.87 times higher than that of NVIDIA’s H2O, exclusive for China as we said. Huawei’s data cannot be verified, since Hopper-era cards do not support FP4 natively, while the Atlas 350 is the first domestically manufactured Chinese accelerator optimized for precision FP4.
That’s already a significant achievement, as even NVIDIA recently started supporting this format with its Blackwell GPUs. FP4 allows larger models to be deployed on the same hardware, requiring less memory.
Speaking of memory, the Atlas 350 includes 112GB of Huawei’s proprietary HBM memory, known as “HiBL 1.0.” Its bandwidth reaches 1.4 TB/s and access granularity has been reduced from 512 bytes to just 128 bytes. Additionally, it supports 2 TB/s interconnect bandwidth using the new LingQu protocol, 2.5 times higher than the previous Ascend 910 series. The Atlas 350 has a power rating of 600 W, 200 watts more than the H2O.
These specifications offer a impressive image for a nationally manufactured chipespecially considering that it is produced under US sanctions. For example, Huawei does not have access to the TSMC foundrie’s CoWoS technology that NVIDIA uses to stack HBM memory close to the GPU, so the company is using other advanced packaging techniques. The memory itself is manufactured in-house and is supposed to compete with that of SK Hynix and Micron, although we do not know who the supplier is.
The exact availability has not been announced, but Huawei kept its promise to announce this Atlas 350 (Ascend 950PR) in the first quarter of 2026. Prices are advanced at around 111,000 yuan (approximately $16,000). A competitive Price compared to NVIDIA’s H2O, whose price in the region ranges between $15,000 and $25,000.
