Science fiction is littered with fantastic visions of computing. One of the more pervasive is the idea that one day computers will run on light. After all, what’s faster than the speed of light?
But it turns out Star Trek’s glowing circuit boards might be closer to reality than you think, Ayar Labs CTO Mark Wade tells The Register. While fiber optic communications have been around for half a century, we’ve only recently started applying the technology at the board level. Despite this, Wade expects, within the next decade, optical waveguides will begin supplanting the copper traces on PCBs as shipments of optical I/O products take off.
Driving this transition are a number of factors and emerging technologies that demand ever-higher bandwidths across longer distances without sacrificing on latency or power.
If this sounds familiar, these are the same challenges that drove telecommunication giants like Bell to replace thousands of thousands of copper telephone cables with fiber optics in the 1970s.
As a general rule, the higher the bandwidth, the shorter the distance it can travel without the assistance of amplifiers or repeaters to extend the reach at the expense of latency. And this is hardly unique to telecommunications networks.
The same laws of physics apply to interconnect technologies like PCIe. As it doubles its effective bandwidth with each subsequent generation, the physical distance the signal can travel shrinks.
“In a lot of cases, long distances are now defined as anything more than a few meters,” Wade said. “As the PCIe bandwidths are going higher and higher, you can no longer escape the server board without putting a retimer on the board” to increase the signal.
“Even if you can get the bandwidth from point A to point B, the question is with how much power and with how much latency,” he adds.
Ayar Lab’s something something something
This is exactly the problem that Ayar Labs is trying to solve. The silicon photonics startup has developed a chiplet that takes electrical signals from chips and converts them into a high-bandwidth optical signal.
And because the technology uses chiplet architecture, it’s intended to be packaged alongside compute tiles from other chipmakers using open standards like the Universal Chiplet Interconnect Express (UCI-express), which is currently in development.
The underlying technology has helped the company raise nearly $200 million from tech giants like Intel and Nvidia, and secure several high-profile partnerships, including one to bring optical I/O capabilities to Hewlett Packard Enterprise’s high-performance Slingshot interconnect fabric.
While Wade firmly believes that optical communication at the system level is inevitable, he notes there are several applications for optical interconnects in the near term. These include high-performance computing and composable infrastructure.
“Our claim is that the electrical I/O problem is going to become so severe that computing applications are going to start to get throttled by their ability to shift bandwidth around,” he said. “For us, that’s AI and machine learning scale out.”
These HPC environments often require specialized interconnect technologies to avoid bottlenecks. Nvidia’s NVLink is one example. It enables high-speed communication between up to four GPUs.
Another area of opportunity for optical I/O, Wade says, is the kind of rack-level composable infrastructure promised by Compute Express Link’s (CXL) latest specs.
CXL defines a common, cache-coherent interface based on PCIe for interconnecting CPUs, memory, accelerators, and other peripherals
The CXL 1.0 and CXL 2.0 specs promise to unlock a variety of memory pooling and tiered memory functionality. However, the open standard’s third iteration, expected to be ratified later this year, will extend these capabilities beyond the rack level.
It’s at this stage of CXL’s development that Wade says optical’s advantages will be on full display.
“Even at the CXL 2.0 level, you’re very limited to the degree in which you can scale out, because the moment you hit something like a retimer, you start to incur latencies,” that make memory pooling impractical, he said.
However, for at least the first generation of CXL products, Wade expects most, if not all, will be electrical. “There’s a lot of software stack work that has to get done to really enable these kind of disaggregated systems” before CXL will be ready for optical I/O, he said.
But as the applications for optical I/O become more prevalent, Wade predicts the supply chain economics will make the technology even more attractive from a cost perspective. “It’s our belief that we’re gonna see an optical I/O transformation start to hit throughout almost every vertical market that’s building computing systems.”
Of course, getting there won’t be without its challenges, and one of the biggest customers is convincing the technology is not only more performant and economically viable but mature enough for production environments.
This is specifically why Ayar Labs is focused on optical interconnects as opposed to co-packaged optics. One of the reasons that co-packaged optics haven’t taken off is their splash radius in the event of failure is significantly larger. If the optics fail on a co-packaged optical switch, the entire appliance goes down. And many of these same concerns apply to optical I/O.
“Whenever you have a heavily commoditized, standardized, risk-averse application space, that is not a place to try to deploy a new technology,” Wade said. However, “if you have a high-value application that highly benefits from increases in hardware performance, then you’re obviously going to take more risk.”
By focusing its attention on HPC environments, Ayar believes it can refine its designs and establish a supply chain for components, all while racking up the substantial field-operating hours necessary to sell to more mainstream markets.
Sci-Fi optical computers still more than a decade away
For customers that are ready and willing to risk deploying nascent technologies, optical I/O is already here.
“The customer that we’re delivering to right now has already replaced their board-level links with our optical I/O,” Wade said. “Every socket-to-socket link is an optical I/O link, and that’s even at the board level.”
As the technology matures, the question then becomes whether the optical waveguides will ever get integrated into the PCB — ala Star Trek.
“Will we see the optical waveguides getting integrated into the boards? I do think we’ll see some of that actually emerge within the next decade,” he said. “As the volume of optical I/O solutions start to get massive, it’ll make it more attractive for some of these solutions.”
Once you start shrinking beyond the board level, the future of optical I/O gets a bit murkier. The next logical step, Wade says, would be using optics to connect the individual dies that make up the chip.
However, he doesn’t expect this to happen anytime soon. “As you go into the millimeter scale, electrical I/O has, I think, a healthy roadmap in front of it,” he said. “Beyond 10-15 years, we might see… optical communication start to enter the millimeter scale regime.” ®