The past few years has seen a growing interest in a switching technology called co-packaged optics, which promises substantial power and cost savings over conventional switch and pluggable systems. And while the first co-packaged optical switches are slated to launch later this year, and chipmakers are investing billions to advance silicon photonics technologies, don’t expect them to be competitive any time soon, according to Kevin Deierling, SVP of networking at Nvidia.

Engineers have been surprisingly successful at putting off co-packaged optics almost indefinitely, he said. “I think I heard for the first time that when we hit 25 Gb/s per link that we’d have to go to onboard optics. Then it was 50 Gb/s, and then it was at 112 Gb/s.”

This week, Deierling’s team announced a 51.2 Tb/s switch platform that fits in a 1U or 2U footprint and delivers up to 64, 800 Gb/s ports or 128, 400 Gb/s ports using 112 Gb/s serializer/deserializers.

“Somehow or another all of the different disciplines just continue to make it work,” he said.

And Deierling doesn’t expect that to change, at least not in the near future. “There’s plenty of really interesting engineering that we can do with chip connections, without having to resort to esoteric onboard optics,“ he added.

Co-packaged Optics Are Inevitable

Deierling isn’t, however, saying there’s no future in co-packaged optics. Eventually, he said, engineers will reach a point where onboard optics can no longer be avoided. “The question is when, and frankly, we don’t see the end” of traditional switching any time soon.

The major problem co-packaged optics promises to solve isn’t port performance, it’s power consumption, heat dissipation, and by extension port density.

“The problem we’re trying to solve with co-packaged optics is power consumption and density … i.e. how many pluggables can you fit on a switch faceplate, and what’s the power consumption for that,” Dell’Oro Group analyst Sameh Boujelbene told SDxCentral.

Today, the business case for co-packaged optics isn’t there and the power envelope and heat dissipation of copper optics is still higher, she said.

“It’s just too early to have a co-packaged optics solution that is ready for mass deployment and volume production within the next few years,” Boujelbene said, adding that 51.2 Tb/s and even next-generation 102.4 Tb/s switches are unlikely to push vendors toward on-board optics.

“For co-packaged optics to make sense, they have to bring significant savings in terms of power consumption,” she said.

Many in the industry are confident they’ll be able to deliver pluggable optics up to 3.2 Tb/s per port without resorting to co-packaged optics, she added.

Much of Deierlings pragmatism around co-packaged optics come down to this problem. “It’s really about building things cost effectively, power efficiently, and that you can yield and manufacture in volume,” he said.

Co-Packaged Optical Challenges

The problem with co-packaged optics, however, goes beyond economics, Deierling said.

The technology is incredibly complex and for all the pros it promises, co-packaged optics isn’t without its challenges.

“There's a lot of downsides in terms of the blast radius of optics,” Deierling said. “That’s one of the least reliable components in a system.”

By blast radius, he's is referring to co-packaged optics potential for failure compared to traditional switch and pluggable combos.

Today, if an optical module fails, only a single port on the switch goes down, and bringing it back online is as trivial as replacing the module, he explained. “When you do onboard optics, bad things can happen in terms of your blast radius, depending on what the failure mechanism is.”

Because the optical transceivers are integrated directly into the silicon, a failure has the potential to take out more than one port. It could take out the entire switch.

Chipmakers Invest Big in Co-Packaged Optics

These challenges haven’t stopped chipmakers from investing billions in co-packaged optics and silicon photonics technologies.

Late last year, Intel Labs announced a new research center to drive faster, more efficient compute interfaces by substituting electricity for light. Meanwhile, Marvell acquired U.S. chipmaker Inphi, in a deal valued at $10 billion.

Inphi specializes in optoelectrical interconnects, commonly used in cloud data centers, where the company can provide up to 800 Gb/s of throughput. The company, in early 2021, teamed up with networking giant Cisco to develop a 51.2 Tb/s co-packaged optical switch, which is expected to launch sometime in early 2024.

A few weeks later, Broadcom unveiled its first co-packaged optical switch, which combined a 25.6 Tb/s Tomahawk 4 switch ASIC with integrated optical interconnects. Dubbed Humboldt, the switch is slated for release in 2022, with a 51.2 Tb/s-capable switch coming in 2023.

So why invest so heavily in a technology unlikely to be competitive anytime soon? Because co-packaged optics are hard, Boujelbene said.

“The reason the industry is talking so much about co-packaged optics is because the learning curve is so steep. There’s so much to solve… and we know at some point we’ll have to go co-packaged,” she said.