Huawei introduces an optical-interconnect Ascend 960 SuperPoD

The optical connection becomes part of the product

Huawei introduced a new SuperPoD at its September 17 Connect conference, naming it Atlas 960E in its English announcement and Ascend 960 in the Chinese version. The central change is near-packaged optics, or NPO: optical components sit closer to the electronics, reducing the distance high-speed electrical signals have to travel.

Huawei's Hi-ONE optical engine is specified at 7.2 terabits per second. TechNode's conference account reports a design supporting up to 4,096 NPU cards, eight exaflops of FP8 computation and one petabyte of high-bandwidth memory. These figures describe the announced system rather than an independently measured installation.

Why a larger cluster needs a different connection

When a model is distributed across many accelerators, those devices repeatedly exchange intermediate results. More chips add arithmetic capacity, but they also create communication work. If a processor waits for data, its theoretical peak cannot be fully used.

Optical links are one way to move that data over a larger system. Bringing optics closer to chips can change power consumption, packaging and cable density. It also introduces practical questions about cooling, serviceability and how a failed component is replaced. Those questions belong in the same evaluation as bandwidth.

The optical design and the software it depends on

Huawei’s detailed release says the fully liquid-cooled Atlas 960E uses 5,500 Hi-ONE optical engines in place of the 48,000 conventional 800G modules in its comparison design. Huawei attributes more than 550 kilowatts of power savings to that change. The comparison concerns the interconnect design; it should not be read as a measured reduction in the entire data center’s electricity bill.

The company also distinguishes one SuperPoD from a network of them. Its announced larger SuperCluster can connect 512,000 NPUs, or as many as one million with an additional multi-rail topology. These are architecture scaling claims rather than a statement that a million-accelerator installation has been commissioned.

The public CANN 9.1.0 documentation provides a useful software baseline. It explicitly adds Ascend 950DT support and names MindSpore, PyTorch and TensorFlow among supported frameworks. It also describes model-specific operators and communication libraries. That release’s 950DT support cannot be treated as proof that every application already works on the announced 960E system.

An overseas development team can use those documents to identify its framework, operator and hardware dependencies before evaluating a deployment. A commercial 960E test still needs an actual supported software release and a disclosed system configuration; the conference announcement supplies neither a universal migration guarantee nor a published international price list.

A system claim needs a system test

Huawei says its NPO design can replace a much larger population of conventional optical modules and reduce power use. It also makes reliability and availability claims. The public launch does not provide an independent test protocol or a like-for-like customer workload comparison supporting those outcomes.

A meaningful comparison would hold the model, precision, output quality and operating conditions constant, then measure useful work per unit of time and energy. Counting accelerators or adding peak arithmetic figures cannot substitute for that exercise.

The broader significance is the product boundary: Huawei is presenting chips, memory, networking and software as one coordinated infrastructure platform. Buyers will need to judge that complete platform, including application migration and maintenance, rather than treating the SuperPoD as a collection of interchangeable accelerator cards.