Huawei's next-gen Ascend NPUs could become China's best option

At Huawei's annual Connect conference, the company unveiled a groundbreaking generation of AI accelerators, named the Ascend 960DT, which promises performance capabilities surpassing those of Nvidia's offerings in China. This new chip, set to launch in the first quarter of 2027, boasts up to 288 GB of custom HiZQ memory technology, half the high-bandwidth memory found in other global devices, and up to four petaFLOPS of FP4 performance.

This represents a significant leap in both performance and memory capacity when compared to Huawei's 950-series chips released earlier this year. When juxtaposed with Nvidia's B300 family of GPUs introduced last year, the 960DT offers similar memory and bandwidth, but only half the FP8 and a third of the FP4 compute. Although a substantial improvement for the Chinese chip designer, there remains a considerable distance to cover before catching up to Nvidia's Rubin and AMD's recently introduced MI455X, both of which promise significantly higher compute and bandwidth.

Importantly, neither of these latter chips is currently available for purchase in China, leaving domestic developers without superior alternatives.