At its annual Huawei Connect event in Shanghai, the company announced major updates to its Ascend chip lineup, marking a new chapter in its AI and infrastructure strategy. Deputy chair Eric Xu described the year as transformative, pointing to breakthroughs like DeepSeek-R1 while acknowledging ongoing challenges in semiconductor manufacturing.
Huawei’s strategy centers on advancing infrastructure design while embracing open-source development. The company is opening up its software stack, including the openPangu foundation AI models and Mind series SDKs, to foster broader adoption.
Expanding the Ascend Portfolio
Three new chip series are on the roadmap: the Ascend 950, 960, and 970.
- Ascend 950: Introduces support for low-precision data formats such as FP8 and MXFP8, delivering performance in the PFLOP range. It also brings improved vector processing and finer-grained memory access.
- Ascend 960: Expected to double the computing power, memory bandwidth, and interconnect capacity of the 950, while introducing Huawei’s HiF4 data format for enhanced precision.
- Ascend 970: Positioned as the flagship, this chip aims for 8 PFLOPs of FP4 performance, 4 TB/s interconnect bandwidth, and significantly expanded memory.
SuperPoDs and the Push Against NVIDIA
Huawei is targeting hyperscalers with its SuperPoD clusters, starting with the Atlas 950 SuperPoD. Equipped with the Ascend 950DT, it will launch ahead of NVIDIA’s NVL144 system and, according to Huawei, deliver over 50 times more NPUs and nearly seven times the processing power. Even against NVIDIA’s upcoming NVL576, Huawei claims its SuperPoD will remain competitive.
General Computing with Kunpeng
Alongside NPUs, Huawei is rolling out two models of its Kunpeng 950 general computing processors. With configurations reaching 192 cores and 384 threads, these CPUs will also form the foundation of the TaiShan 950 SuperPoD—promoted as the world’s first general-purpose computing SuperPoD.
UnifiedBus 2.0: Open Connectivity
A key element of Huawei’s scaling vision is UnifiedBus 2.0, the next generation of its interconnection technology. Unlike its predecessor, the new protocol will be open-source, enabling developers to integrate and scale SuperPods into SuperClusters.
The first such cluster, the Atlas 950 SuperCluster, will outperform today’s largest compute systems, including xAI’s Colossus. By 2027, Huawei expects the Atlas 960 SuperCluster to integrate over a million NPUs and reach 4 ZFLOPS of FP4 performance.
Meeting Surging Compute Demand
Xu framed the company’s direction as a direct response to skyrocketing global demand for compute. With its new Ascend roadmap, open protocols, and large-scale clustering approach, Huawei is positioning itself as a formidable competitor in the high-performance AI infrastructure race.
