Huawei Announces Ascend 950 Cloud Commercial Launch and Accelerated Ascend 960 Roadmap
Huawei announced that its Lingqu Ascend 950 AI compute cluster cloud service will become commercially available in China on September 30 and globally on November 30. Furthermore, Huawei revealed that its next-generation Ascend 960 chip family is ahead of schedule with doubled performance, moving its release up to 2027. The deployment of Ascend 950 clusters and the accelerated roadmap for the 960 series demonstrate Huawei's push to provide a viable domestic AI computing ecosystem alternative to Western hardware. This rapid release cycle aims to support large language model training requiring exascale compute power and unified high-bandwidth memory architectures. The Ascend 950 SuperPOD links 1,024 chips using Huawei's proprietary Lingqu interconnect to deliver 1 EFLOPS of FP8 compute capacity and a 256TB global unified memory space. The upcoming Ascend 960DT will offer 2 PFLOPS FP8 performance, up to 288GB of HBM with 9.6 TB/s bandwidth, and will utilize mass-produced Near Package Optics (NPO) technology.
## BACKGROUND
Huawei's Lingqu protocol (UnifiedBus) is a proprietary high-speed interconnect designed specifically for AI super-nodes to efficiently link processors and memory without relying on traditional interfaces like NVLink or PCIe. Near Package Optics (NPO) places optical communication engines on the board directly adjacent to host logic chips, significantly increasing bandwidth and energy efficiency compared to standard pluggable transceivers.