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

At Huawei's Connect conference, the company revealed its next generation of AI accelerators, the Ascend 960DT, which will likely become China's top choice for AI development. This chip outperforms anything Nvidia currently sells in the region, with up to 288 GB of custom HiZQ memory tech, four petaFLOPS of FP4 performance, and a launch date three quarters ahead of schedule - in the first quarter of 2027.

The Ascend 960DT offers twice the performance and memory capacity of Huawei's 950-series chips, while matching memory and bandwidth to Nvidia's B300 family of chips. While it still trails behind Nvidia's Rubin and AMD's MI455X in terms of compute and bandwidth, the performance gap may not be as significant as it seems because AI models are increasingly scaled across multiple GPUs or NPUs.

Huawei's NPO technology enables it to scale its compute domain to 4,096 chips, capable of delivering up to 16 exaFLOPS of FP4 compute. The company also unveiled a compute-optimized version of the chip, the Ascend 960PR, which will provide up to 8 petaFLOPS of FP4 performance, 192 GB of custom HiBL memory, and 2.4 TB/s of bandwidth.

This heterogeneous architecture, similar to Nvidia's Rubin and Groq LPX configurations, is expected to enhance LLM inference performance by offloading memory-intensive tasks to the 960DT and handing off the compute-heavy prefill phase to the 960PR. Huawei aims to eventually deploy clusters with half a million NPUs, and has already successfully scaled its Ascend 950-based TaiShan superpod to 4,096 accelerators using a two-tier, four-plane Clos network topology.