Huawei’s pitch at its Connect conference in Shanghai this week wasn’t a chip that beats Nvidia’s one for one. It was a way to wire thousands of its Ascend processors together so they behave like a single machine.
The centerpiece is the Atlas 960E SuperPoD, which Huawei calls the industry’s first to use near-packaged optics. One pod holds 4,096 NPUs, delivers 8 exaFLOPS at FP8 precision and carries a petabyte of HBM, The Next Web reported. Huawei’s own optical engine, Hi-ONE, moves 7.2 Tbit/s per unit and has a built-in light source. Fitting 5,500 of them replaces tens of thousands of 800G pluggable optical modules (48,000 by The Next Web’s count, 40,000 in TechNode’s), which Huawei says cuts power draw by more than 550 kilowatts and lifts availability to 99.8%.
A faster chip schedule, a smaller pod
Most attention went to the roadmap. Rotating chairman David Wang said the training-focused Ascend 960DT will arrive in the first quarter of 2027, three quarters ahead of plan, with the inference-oriented 960PR following in the third quarter. He placed an Ascend 970 in 2028 and an Ascend 980 in 2029.
Not every number moved Huawei’s way. China tech analyst Rui Ma pointed out that Huawei had earlier described an Atlas 960 SuperPoD scaling to 15,488 chips, while the pod announced this week holds 4,096. “The chip itself is coming WAY earlier, but the SuperPoD they announced is much smaller than what they originally laid out,” Ma wrote. The Register puts the 960DT at up to 288GB of memory and four petaFLOPS of FP4, roughly half the FP8 and a third of the FP4 compute of Nvidia’s B300, a chip Nvidia can’t sell in China.
UnifiedBus does the heavy lifting
Holding the system together is UnifiedBus. Yang Chaobin, who runs Huawei’s ICT business group, said it folds more than ten interconnect protocols into one, raising bandwidth from hundreds of gigabytes per second to terabytes and cutting latency from seven microseconds to two. In a conventional 100,000-NPU cluster, he said, as little as 20% of compute goes to the model while the rest waits on data movement. Huawei’s lab simulation shows pods of 4,000 NPUs reaching 2.75 times the model FLOPs utilization of eight-NPU servers.
Stacked in a multi-rail topology, Huawei says, the design scales to a million NPUs, though The Register describes that configuration as theoretical. On software, external developers now make up 61% of the community around CANN, Huawei’s counterpart to Nvidia’s CUDA, and Ascend is now an official PyTorch accelerator backend.
Huawei also said more than 1,000 Ascend 910C supernodes have been deployed and Ascend 950 supernodes are in commercial use. Rotating chairman Eric Xu told reporters that demand outstrips supply and the company is prioritizing Chinese customers, with no broad global expansion plan.
WebProNews is an iEntry Publication