Ayar Labs to Demo ‘Commercial-Grade’ 4Tbps Optical I/O Solution

By Oliver Peckham

Ayar Labs, founded in 2015, is pursuing optics as a means of driving higher interconnect speeds and efficiencies in computing. Now, the company is announcing that it plans to demonstrate a staggering 4Tbps optical solution at the Optical Fiber Communication Conference (OFC) in San Diego next week. In an interview with HPCwire, Ayar Labs CEO Charlie Wuischpard called the solution “the full realization of [Ayar Labs’] commercial-grade product” and discussed the future of the company and its products.

But first, some technical details. The optical solution that Ayar Labs will be demoing consists of two key components: a SuperNova laser light source and a TeraPHY optical I/O chiplet. According to Ayar Labs, each SuperNova powers eight fiber links, each of which, in turn, is capable of 256Gbps. That gets you 2.048Tbps in each direction, or 4.096Tbps bidirectional. The connection operates with near-perfect accuracy (“1e-15 errors,” per Wuischpard) and requires just five picojoules of energy per bit. The hardware is primarily manufactured by GlobalFoundries in upstate New York.

As “chiplet” might have clued you in, Ayar Labs’ tech is not designed to supplant system-scale interconnects. “[People] think co-packaged optics means this thing that’s working in an Ethernet or InfiniBand network,” Wuischpard said. “We’re really addressing the compute-to-compute or compute-to-memory communications.” As more accurate comparisons, he cited Nvidia’s NVLink and AMD’s Infinity Fabric.

The pitch for the product is simple: interconnects simply aren’t scaling at the same rate as other system components like compute and memory hardware. “The power consumption to transmit data … is growing at a rate where it’s taking up more power to transmit than it does to compute on that data,” Wuischpard said. And, as Ayar Labs’ front page headline (“Enabling the next phase of Moore’s law through optical connectivity”) makes clear, gains in those other areas are slowing, leaving system-makers and operators looking elsewhere for gains.

Ayar Labs, for its part, is accelerating to meet that need. For the past few years, the company has been exhibiting more elements of – and faster connections via – its optical solution at conferences, but it sounds like this month’s announcement may be the biggest yet. “I think what we’re showing is the full realization of the commercial-grade product,” Wuischpard said, “whereas before we showed subsets of that over time.”

The company has also, slowly but surely, begun shipping its product.

“We started talking about shipments of our product last year and we started shipping our first hundreds of units – this is both chiplets and lasers,” Wuischpard said. “Those are going into prototype machines that our customers are building.”

“This year, because we’ve qualified the manufacturing process, that goes from hundreds into thousands,” he added. “So the revenues are still somewhat modest, but they’re all in a pre-large-scale rollout phase.”

Those thousands of units are going to around a dozen customers, Wuischpard said, up from “a few” last year. While most of those are private, others – like Intel – aren’t, and Wuischpard said Ayar Labs’ investors (which include HPE, Lockheed Martin, Nvidia and others) provide good hints.

“Importantly, it’s not here today yet,” Wuischpard hedged. “We’re still at the front edge – we’re starting to ship in limited volumes for very specialized projects.”

Still, the CEO sees substantial near-term potential for the technology, citing a range in the next two to four years where he expects revenues to significantly ramp up.

“More and more, you’re seeing all the major semiconductor players start to talk about optical I/O as part of their long-term direction,” he said. “All of the exascale or post-exascale initiatives that are underway now for the future – whether it’s the machines in Korea, Japan, Europe or the US – are starting to look at this technology as one of the innovations that can be adapted for those future HPC systems. … Once you establish them and de-risk them at that level, they tend to flow into the broader market.”

Vis-à-vis these larger machines, Wuischpard specifically called out power efficiency as a driving concern.

“Of all the other metrics, power efficiency is one of the biggest drivers of this technology, because if you add it up across an entire system, it turns out to have a massive impact on the power consumption of the cluster or the datacenter,” he said. “I think the first person that puts this in wins Green500 hands-down, no problem.”

Wuischpard is, naturally, confident about the strength of Ayar Labs’ product for these applications. “Relative to competitors, I think there’s nobody even close,” he said. “I think you might see people talk about 400Gbps solutions, 800Gbps solutions, maybe 1.6Tbps solutions – but they all suffer from size and power and latency issues.”

“We think there will be people at this show who start to announce things that look similar – but they probably won’t show the full functionality of their solutions,” he added. Wuischpard repeatedly emphasized that Ayar Labs wanted company in the space, saying that “You don’t want to be the only guy out there” and calling out public demos as part of the company’s strategy (“As an ecosystem kind of play, you don’t want to be a stealth company”).

To that end, Ayar Labs is working to launch a new industry standard for a multi-wavelength light source that Wuischpard said already has 45 different “top players” signed on. And, of course, returning to the headline, Ayar Labs will demo its 4Tbps solution at the conference next week, along with some new developments on packaging and fiber attachment processes (“Attaching fibers into a CPU package is not something that’s been done before”).