Vendors

Huawei believes NPO is the optimal choice for the next generation of optical interconnects for AI computing

Huawei believes NPO is the optimal choice for the next generation of optical interconnects for AI computing

Huawei is betting that a technology called Near-Packaged Optics, or NPO, will become one of the main building blocks of the next generation of AI data centres.

The reason is simple: AI requires massive computing power, but it also depends on high-speed data exchange across increasingly large computing clusters. As AI systems grow from a handful of chips to clusters containing thousands of them, the connections between those chips are becoming a major problem.

Traditional electrical connections struggle when data has to travel fast and over longer distances. They consume more energy, generate heat and can slow down performance. Huawei believes optical connections - which use light rather than electrical signals to move data - are the answer.

But there is a debate over how these optical connections should be built.

One option is Co-Packaged Optics, or CPO. This places optical components extremely close to the main processor inside a switch. In theory, this can deliver very high performance, low delay and lower energy use.

The catch is that it is difficult and expensive to make. It also creates a maintenance headache: if one small part fails, operators may have to replace a much larger and more costly piece of equipment. Huawei argues that this makes CPO less suitable for mass adoption at the moment, even if it could play a bigger role further down the line.

Huawei is instead promoting NPO as a more practical middle ground.

With NPO, the optical components sit close to the main chip, but remain separate and replaceable. That means the system can still deliver faster speeds and use less power than today’s conventional pluggable optical modules, while avoiding some of the complexity of CPO.

Huawei says NPO can also be made using existing supply chains, helping manufacturers bring products to market more quickly. This matters because the AI industry is moving fast, with cloud companies and chipmakers racing to build larger and more powerful systems.

However, technology alone will not be enough. Different companies need their equipment to work together, which means agreeing on common standards.

Huawei has been pushing for this through the Optical Internetworking Forum, or OIF, an international industry body. In May, the organisation approved a project to develop standards for 12.8-terabit-per-second NPO modules. More than 40 companies are involved, including cloud providers, equipment makers and component suppliers.

The standards are expected to cover the physical design of the modules, power, cooling, connectors and the way they communicate with other equipment. Huawei says this is important because it could prevent the market from splintering into incompatible systems built by rival vendors. Huawei says the proposed OIF NPO standard has drawn participation from industry players across the United States, Canada, China and Japan. The company expects the effort to result in an open, globally applicable standard, potentially enabling greater interoperability and broader participation across the optical interconnect ecosystem.

The point is similar to Bluetooth. A Bluetooth mouse can connect to laptops made by different companies because the devices follow a common standard. Huawei argues that standardising NPO would create the same kind of interoperability, allowing parts from different suppliers to work together more easily.

To show that NPO is ready to move beyond the drawing board, Huawei has introduced a 7.2-terabit-per-second NPO module. It says the new module can use around 60 per cent less power and deliver far lower delay than conventional alternatives.

Huawei’s message is that while there might be multiple technology paths for the future technology evolution, NPO offers a better balance of speed, cost, energy efficiency and ease of maintenance - making it the more realistic option for the AI boom now underway.



More Articles you may be Interested in...