Huawei unveils new Atlas 350 AI accelerator with 1.56 PFLOPS of FP4 compute and up to 112GB of HBM — claims 2.8x more performance than Nvidia's H20

Wait 5 sec.

The Atlas 350 is using the Ascend 950PR chip, but it looks like a cut-down version with less compute. Translation overhead aside, previous reports and announcements have already revealed the full specs of this silicon, such as 128 GB of HBM and up to 2 PFLOPS of FP4 compute, but the Atlas 350 is being reported just a smidge below those numbers.