Huawei Ascend 960 SuperPoD aims to surpass Nvidia AI hardware in China — 960E with over 4,000 NPUs and an upgraded UnifiedBus Interconnect aims for 10 trillion parameters

Huawei's Ascend 960 roadmap aims to outperform Nvidia's AI hardware by scaling up memory capacity and interconnect speeds. The new Atlas 960 SuperPoD features up to 15,488 processors with 34 PB/s bandwidth, setting it apart from earlier configurations. This system enables 30 EFLOPS of FP8 computing and 60 EFLOPS using FP4, alongside a massive 4,460 TB of memory. The interconnect bandwidth is specified at 34 PB/s, allowing for 15.9 million tokens per second of training performance.

Huawei's UnifiedBus interconnect architecture links processors, memory, and storage, reducing communication bottlenecks within AI systems. The 960E variant of UnifiedBus connects up to 4,096 NPUs, providing shared memory access across the SuperPoD. By combining UnifiedBus with Hi-ONE optical technology, Huawei reduces power consumption and improves transmission speeds compared to traditional optical connections.

The Atlas 960E system can utilize approximately 5,500 Hi-ONE units, which would typically require about 48,000 conventional 800G optical modules. This configuration not only reduces power consumption by over 550 kilowatts but also allows the system to support models containing up to 10 trillion parameters. Huawei's strategy combines thousands of processors with faster communication, enabling continuous data exchange during demanding AI training and inference workloads.