Huawei has disclosed the Atlas 960E SuperPoD, an AI compute cluster node that integrates 4,096 NPUs and delivers 8 EFLOPS of FP8 compute — and it is the first such system to pair NPO (near-package optics) with Huawei's Hi-ONE optical interconnect architecture.

The headline figures set the scale. Four thousand and ninety-six NPUs operate as a single pod. Aggregate throughput reaches 8 exaFLOPS at FP8 precision, the low-precision format now dominant in large-model training and inference. No previous Huawei system has combined this NPU count with a fully optical node-level fabric.

What does the SuperPoD change for AI infrastructure?

The Atlas 960E SuperPoD shifts the interconnect bottleneck from copper to optics at node scale. NPO — near-package optics — places optical engines close to the compute silicon rather than at the board edge, shortening electrical paths and cutting the power and latency penalties that large copper fabrics impose.

Hi-ONE is Huawei's name for the optical network architecture that ties the pod together. In a 4,096-NPU pod, the fabric must carry traffic between all accelerators with minimal hops. An optical approach allows Huawei to scale the pod as one logical unit instead of linking smaller racks with conventional networking.

The practical consequences for operators:

• Higher effective utilization. Faster interconnect reduces the idle time GPUs and NPUs spend waiting on collective operations such as all-reduce.
• Lower power overhead per delivered FLOP. Optical links consume less power than high-speed copper at equivalent reach.
• Simplified pod topology. A single SuperPoD replaces multi-tier switch hierarchies that would otherwise connect thousands of accelerators.

Why does FP8 throughput matter?

Huawei quotes the 8 EFLOPS figure in FP8, not FP16 or FP32. This mirrors where the industry has moved. Modern large language model training runs increasingly use 8-bit floating point for matrix math, halving memory and bandwidth requirements relative to FP16 while preserving training stability through careful scaling.

At 8 EFLOPS FP8, one Atlas 960E SuperPoD sits among the largest single-node AI compute units disclosed publicly. The FP8 rating is the number that matters for workload planning, because it reflects the throughput available to the frameworks and precisions customers actually deploy.

What is the significance of "first NPO/Hi-ONE"?

The claim of first matters for two reasons.

First, NPO is an emerging packaging technique. Most current accelerator clusters rely on pluggable optical modules or direct-attached copper. Moving optics near the package increases bandwidth density and signal integrity, but it is harder to manufacture and service. A shipping system built around NPO signals that the technology has cleared reliability and cost thresholds at production volume.

Second, Hi-ONE as an integrated architecture means Huawei controls the full stack — NPU silicon, optical interconnect, and pod-level software. For buyers facing export controls and supply restrictions on Western accelerators, a domestically designed pod with an optical fabric from a single vendor offers a coherent alternative to assembling clusters from components sourced across multiple suppliers.

Who is the Atlas 960E SuperPoD for?

A 4,096-NPU pod with 8 EFLOPS FP8 targets customers training foundation models or running hyperscale inference. The system functions as the building block for larger data center deployments: multiple SuperPoDs can in principle connect into clusters exceeding the single pod's capacity.

The disclosure positions Huawei in direct competition with the largest Western superpod designs, where 4,096-accelerator units have become the reference granularity for frontier-scale training.

What comes next?

Pandaily, which first reported the specification, lists the Atlas 960E SuperPoD's NPO/Hi-ONE integration, 4,096-NPU count, and 8 EFLOPS FP8 rating as confirmed. Huawei has not yet published pricing or general availability dates in the reported material. Buyers evaluating the platform will need those figures, plus independent benchmark results, before committing workloads.

For now, the numbers stand on their own: 4,096 NPUs, 8 EFLOPS FP8, and the industry's first node-level pairing of NPO optics with the Hi-ONE fabric.