Huawei has announced what it calls the, “world’s most powerful SuperPoDs and superclusters”.
In AI nomenclature, a SuperPOD is a group of racks that work as a single unit integrating compute, network, software, storage and management. A supercluster is a group of superPODs.
Given that Nvidia trademarked SuperPOD, the Chinese company was smart enough to call its product SuperPoD (AMD has its own version called MegaPOD). Its Atlas 950 SuperPoD comprises of 8,192 Ascend NPU (essentially AI accelerators) with a superior version, the 960, delivering almost twice that number at 15,488 (89% more).
A supercluster with 1-million NPUs
The Atlas 950 will use the newly announced Ascend 950 chips while the Ascend 960 series will fit in the Atlas 960. The 950 series will be available in Q1 26, while the 960 will come in Q4 27 and – you’ve guessed it – there’s an Ascend 970 planned for Q4 2028.
Huawei’s deputy chairman, Eric Xu, went on to claim that its SuperPoDs “are currently the most powerful SuperPoDs in the world, and will remain so for years to come”, based on publicly available roadmaps from Nvidia and AMD (although he didn’t name them).
Xu went on to announce superclusters based on the Atlas 950 and Atlas 960, with more than 520,000 (64 superPoDs) and more than one million NPUs, respectively (at least 66 superPoDs).
This, Xu posits, will outstrip xAI Colossus, currently the world’s largest computing cluster.
The caveat is that xAI is already working on Colossus 2, one of a number of Gigawatt-class clusters that will almost certainly surpass 950-based and 960-based superclusters. Oracle/OpenAI, Meta and AWS/Anthropic are also building such (hyper)clusters as well.
A third surprise announcement was the launch of UnifiedBus, which is Huawei’s alternative to Nvidia’s Infiniband. The company is keen to create an open UnifiedBus ecosystem, but its press release doesn’t mention whether this interconnect protocol will be open-sourced.
Huawei’s claims – if true – are impressive: 100x improved reliability for optical interconnect, with maximum range extended to more than 200 meters (almost 700 feet) and NPU-NPU latency reduced to 2.1ms, a 30% improvement over current technologies.
Chinese companies and China’s central government are most certainly going to be the main customers of Huawei’s SuperPoDs and superclusters for now.
Pricing, performance and power consumption are yet to be detailed.
You might also like
https://cdn.mos.cms.futurecdn.net/osqAuJh7DcU8cC2GPiRFKM.jpg
Source link
desire.athow@futurenet.com (Desire Athow)
