• 2 min read
Huawei debuts Atlas 950 with 256TB memory and 1 EFLOPS FP8
Huawei says its Atlas 950 SuperPoD packs 1,024 Ascend chips, 256TB of unified memory, and up to 6.7x the compute of Nvidia NVL144.
Image: iXBT
Huawei has unveiled the Atlas 950 SuperPoD, a new supercomputer the company describes as the industry’s largest system for AI computing. According to Huawei, the platform combines 1,024 Ascend chips and can scale to connect up to 8,192 NPUs.
SuperNode, or SuperPoD, systems are large-scale AI computing platforms that tie together thousands of AI chips — including GPUs and NPUs — into a single logical unit using high-speed interconnects. The goal is to improve the efficiency of training large AI models.
Huawei says the Atlas 950 delivers terabyte-level NPU bandwidth and 3 microseconds of ultra-low latency. The company used its own Lingqu protocol and an upgraded SuperPoD architecture for the new AI cluster.
Huawei says the system is designed to handle training and inference for trillion-parameter, multi-level models. The company also claimed the Atlas 950 SuperPoD can deliver 6.7 times more computing power and 15 times more memory than Nvidia NVL144, a server rack based on 144 Rubin-series accelerators using the Rubin Ultra architecture for AI training.
Recommended reading
FBI eyes Nvidia and Google TPU AI systems
On raw specs, the Atlas 950 offers 1 EFLOPS FP8 and 2 EFLOPS FP4 of compute, paired with 256TB of unified memory.
Enterprise Editor
Marcus follows the money. He covers enterprise software, cloud architecture, and the tectonic shifts in Big Tech strategy. He translates dense earnings calls and complex M&A activity into actionable insights about where the industry is actually heading. If a tech giant makes a silent pivot, Marcus is usually the first to notice.
via iXBT
