Huawei said it is testing an Atlas 960 SuperPoD AI compute unit that will for the first time use near-packaged optics (NPOs) to help boost performance and reduce latency.
The system, which also uses Huawei’s UnifiedBus interconnect technology, is part of a strategy of building systems using massive groupings of chips that act as a single computer.
Huawei said the use of UnifiedBus and the Hi-ONE optical system would allow a SuperPoD to connect up to 4,000 processors.
Compute cluster
The Atlas 960 SuperPoD, using Huawei’s Ascend 960 chips, is expected to deliver 2.3-fold and 2.5-fold improvement in training and inference for 10 trillion-parameter models, compared to the Atlas 950 SuperPoD, which rolled out commercially last year.
The company said last year that 64 Atlas 950 SuperPoDs can be combined into an Atlas 950 SuperCluster to deliver 524 exaflops of FP8 compute power.
More than 1,000 units across the SuperPoD line have been commercially deployed, Huawei said at the conference.
The Atlas 950 SuperPoD runs 8,192 Ascend AI chip, while the Atlas 960 SuperPoD expands the scale to run up to 15,488 Ascend cards across 220 cabinets covering 2,200 square metres, according to Huawei.
At the conference, Huawei said it is bringing forward the target launch for its Ascend 960 DT AI training chip by three quarters, to the first quarter of next year.
The launch of the Ascend 960 PR chip for AI inference has been brought forward by one quarter to the third quarter of 2027.
The clustering strategy is in part a way of compensating for China’s lack of high-end chip manufacturing technology, but Huawei says it began working on the technology prior to the US sanctions that have banned sales of such manufacturing tools to China, and believes the broader industry will adopt a similar approach.
Huawei unveiled its first AI chip strategy in 2018, and began increasing its investments in computing architectures and interconnect technologies after launching its SuperPoD and SuperCluster products in 2019.
High-speed link
“In the years to come, other players would certainly follow suit,” Huawei rotating chair Eric Xu Zhijun told the .
Huawei’s UnifiedBus competes with Nvidia’s proprietary NVLink architecture.
A coalition of companies including AMD, Intel, Google, Microsoft, Meta, AWS and Alibaba is backing the rival Ultra Accelerator Link, or UALink, as an open system to counter Nvidia’s dominance.