loading
loading
Huawei has built its 7.2T Hi-ONE near-packaged optical engine into the new Atlas 960E SuperPoD. NPO has moved remarkably quickly from an architectural argument to standards work, functioning hardware and now an AI computing system. Near-packaged optics has just passed an important test. Not a laboratory test. A relevance test. Only a few weeks ago, the argument around NPO was largely architectural. Could moving the optical engine closer to the processor deliver enough of the power and bandwidth advantages of co-packaged optics without inheriting all of its manufacturing and maintenance problems? Then came the standards discussion. Huawei demonstrated a 7.2 Tbit/s NPO engine at CIOE in Shenzhen while the wider industry began working through the much harder problem of creating common interfaces around the technology. Now Huawei has put NPO into a computing system. At Huawei Connect 2026 in Shanghai, the company unveiled the Atlas 960E SuperPoD, built using its Hi-ONE 7.2T optical engine and UnifiedBus interconnect architecture. This is where the story gets interesting. 5,500 optical engines replace 48,000 modules A single Atlas 960E can scale to 4,096 NPUs, according to Huawei, delivering 8 EFLOPS at FP8 precision and up to one petabyte of high-bandwidth memory. The optical numbers are more revealing. Huawei says the system uses 5,500 Hi-ONE engines instead of the approximately 48,000 conventional 800G optical modules that would otherwise be required to connect its NPUs. The company claims that cuts power consumption by more than 550 kilowatts, doubles fault-free operating time and produces system availability of 99.8%. Just consider the scale of the proposed change. AI infrastructure is already running into limits involving power, cooling, space and the efficiency with which expensive processors can communicate. Removing tens of thousands of individual optical modules from a large computing system is therefore not some marginal component optimization. If Huawei’s claims hold up, it changes the economics and engineering of the system around the chips. The interconnect problem This is the issue that has been creeping up on the AI industry while most public attention remains fixed on processors. A cluster containing more GPUs or NPUs is not automatically a better computer. They have to communicate. Huawei says communications inside conventional 100,000-NPU clusters can consume more than 40% of training time. Its own Markov Lab simulations suggest a cluster organized around 4,000-NPU SuperPoDs could deliver 2.75 times the Model FLOPs Utilization of one constructed from conventional eight-NPU servers. The underlying problem is not controversial, however. As AI systems grow larger, the performance of the links between processors increasingly determines how much useful work the processors themselves can perform. Huawei’s answer is the SuperPoD. Rather than treating thousands of processors as separate servers connected by a network, it is attempting to bind them together closely enough that they behave more like one very large computer. That requires memory, compute and networking to operate as parts of the same architecture, which is where UnifiedBus and NPO come in. Hi-ONE gets a job Hi-ONE is Huawei’s 7.2 Tbit/s near-packaged optical engine. The company says it is the first NPO product ready for mass production and the first with an integrated light source. The significance of Atlas 960E is that Hi-ONE now has somewhere to go. At CIOE, Huawei demonstrated that the 7.2T engine worked. Atlas 960E shows what Huawei wants to do with it. The optical engines sit within an architecture based on UnifiedBus, Huawei’s attempt to create a common high-speed interconnect across the computing system. Unified memory addressing allows processors across physical nodes to operate as part of the same logical machine. That is a more consequential use of optics than simply moving traffic between racks. The network is becoming part of how the computer computes. This is also why the power numbers deserve attention. The AI infrastructure problem is no longer just how much electricity the processors consume. Every watt spent moving information between those processors becomes part of the cost of producing intelligence. Saving hundreds of kilowatts in the interconnect of a single system starts to matter very quickly when those systems are deployed at scale. Huawei wants to go much bigger Atlas 960E is not the end of Huawei’s architecture. The company says multiple SuperPoDs can be connected through UnifiedBus into much larger SuperClusters. Its proposed two-tier, four-plane Clos architecture can connect as many as 512,000 NPUs. Huawei says a multi-rail configuration could eventually take that figure to one million. Numbers at that scale should be treated carefully. Designing an architecture capable of addressing one million processors is not the same thing as proving that a million-processor system will deliver useful performance economically and reliably. Still, Huawei’s direction is unmistakable. It is trying to compete in AI not simply by producing a faster accelerator, but by engineering the entire computing system around it: processors, memory, storage and increasingly the optical interconnect. For an optical industry accustomed to describing itself in speeds, feeds and transmission distances, that represents quite a promotion. What happened to the NPO argument? NPO began as an answer to an awkward engineering question. Traditional pluggable optics puts the optical module far enough from the processor that increasingly fast electrical signals must travel across the circuit board to reach it. As speeds increase, that journey becomes more difficult and consumes more power. Co-packaged optics solves much of that problem by putting optics beside the switching silicon, but introduces difficult questions around packaging, thermal management, manufacturing and maintenance. NPO sits between the two. That argument was the subject of the first article in this series. The second examined whether NPO can develop the standards and interoperability required to become a genuine multi-vendor market. Huawei’s launch of Atlas 960E makes the standards question more important. If near-packaged optics becomes embedded deeply inside AI computing systems, customers will have even more reason to demand common electrical, optical, mechanical and management interfaces rather than become dependent on proprietary implementations. The Optical Internetworking Forum has already begun that work, with Huawei among the companies participating in its NPO efforts. The technology and the ecosystem now have to catch up with each other. NPO has moved very fast This is what makes the past few weeks notable. NPO was an architectural proposition. Then it became a standards project. Then Huawei demonstrated 7.2T hardware. Now that hardware sits inside a SuperPoD designed to connect thousands of AI processors. Just like any new technology innovation, for NPO, there are still questions to be answered. Huawei’s power, reliability and performance claims need independent verification. The interoperability work is unfinished. Competing implementations will emerge. CPO is not going away. But NPO no longer has to prove that somebody can find a use for it. Huawei just did.
document fetched 2026-09-24 02:51 · canonical https://fierce-network.com/sponsored/huawei-takes-npo-lab-ai-machine