Huawei Unveils SuperPoD Computing Portfolio

The Atlas 950 SuperPoD architecture clusters 8,192 of Huawei's new Ascend 950 NPUs (Neural Processing Units). The underlying Ascend 950 chip supports new low-precision formats such as MXFP4 and the proprietary HiF8, with a 2 TB/s interconnect bandwidth. Alongside the AI-focused Atlas hardware, the portfolio includes the TaiShan 950 SuperPoD for general-purpose computing. This system is built on the forthcoming Kunpeng 950 processor, which will be offered in configurations up to 192 cores and 384 threads, and is positioned to compete with legacy mainframe and Oracle Exadata systems. Huawei claims a single Atlas 950 SuperPoD can deliver 8 FP8 ExaFLOPS, directly challenging Nvidia's roadmap by comparing it to the 1.2 FP8 ExaFLOPS of the forthcoming Vera Rubin NVL144 platform. The Atlas 950 SuperCluster, a larger configuration, aims for 524 FP8 ExaFLOPS for training performance. However, this performance comes with a significant physical footprint. The Atlas 950 SuperPoD requires 160 total cabinets spread across 1,000 square meters. In contrast, analysis suggests an equivalent amount of compute from Nvidia's next-generation hardware may require as few as five racks, implying a 32x smaller footprint for competitors. To connect these large clusters, Huawei has developed its own proprietary interconnect, UnifiedBus 2.0, an open-source protocol designed to compete with technologies like Nvidia's NVLink. This enables the scaling of multiple SuperPoDs into a single SuperCluster, with the Atlas 960 generation aiming to connect over one million NPUs. Crucially, the entire Ascend chip family faces severe regulatory hurdles. The U.S. Commerce