Notes From OCP: Agnostic on the Package, Practical About the Laser

11 minutes ago

4 min read

I spent a week at the OCP Global Summit in San Jose this month, and almost every conversation I had with analysts and companies building optical links came back to lasers. In fact, several companies came to us asking for them. After all these conversations, here are my thoughts on the current state of play.

By Matt Crowley, CEO of SCINTIL Photonics

NPO or CPO: the package is the system builder’s call

There was a lot of discussion about near-packaged optics (NPO), where the optical engine sits on the board close to the processor, and co-packaged optics (CPO), where it shares the processor’s package. The timing and standards picture is still unsettled. There doesn’t yet seem to be an NPO standard, although the Open CPX MSA is now developing specifications that cover two socket types. On CPO, most people I spoke with expect it to happen; the debate is whether that’s 2028 or 2030.

Scintil’s position is that we’re agnostic. Our technology is perfect for both. The transceiver builder decides whether the laser goes inside the package or outside it, for example, in an ELSFP, the faceplate-pluggable laser module the OIF has specified. They’ll decide based on density, thermals, and serviceability, and the light source should fit either way.

The laser shortage is starting to shape architecture

The supply picture behind all those conversations is real, and it’s getting tighter. In early October, Lumentum, one of the main suppliers of the indium phosphide lasers these links depend on, told Bloomberg that its optical components are “completely sold out” through early 2029 and that it can’t meet about 70% of demand for some products through 2027. Further up the chain, DigiTimes reported in September that the silicon photonics boom could push the supply-demand gap for indium phosphide, the material these lasers are made from, above 70%. Multiwavelength laser sources are of course much more demanding of manufacturing precision than single-wavelength sources, so the industry will face even bigger challenges in laser sourcing in the future.

When lasers can’t be delivered in volume, designers start looking at other architectures. I heard a lot of talk at the summit about VCSEL and micro-LED approaches, along with fair skepticism about them. One concern I heard was the cost of the multicore fiber; one demo I saw had about a thousand fibers in a single cable bundle. I also don't see how those approaches would work with the upcoming deployment of optical circuit switches in the scale-up network.

The architecture that wins is the one you can manufacture

New interconnect ideas go through three tests. A system architect proposes something, the photonics engineers check whether it’s physically possible, and then the supply chain has to show it can build the parts in volume. Some ideas stop at the second test. Others get past it and stop at the third; however elegant an architecture is, the one that wins is the one you can manufacture.

We built SHIP™, our heterogeneous integration process, around that third test. SHIP™ adds lasers to a standard silicon photonics flow. The gratings that set each laser’s wavelength are patterned in the silicon by deep-UV lithography, which foundries already do precisely and repeatably, and the indium phosphide gain material is bonded only where a laser will be. SHIP™ is validated on Tower Semiconductor’s silicon photonics production lines, and the first tens of thousands of units are in process at the fab. LEAF Light™, our single-chip laser source for dense wavelength-division multiplexing (DWDM), is the first product built on SHIP™, and we've sampled it to customers.

More colors = low power

DWDM carries many colors of light on a single fiber, each color a separate data channel, giving designers a second way to scale. Rather than pushing each channel to run faster, you can add more channels and run each one slower. Because the electronics that drive a slower channel use less power, power per bit doesn't climb as bandwidth grows.

Once you can put many wavelength lasers on a single chip, multiwavelength is the best architecture because it lets you add bandwidth while improving power efficiency. LEAF Light™ has demonstrated up to 16 lasers on a single chip, with up to 16 colors spaced as close as 100 gigahertz (GHz) apart with ±10 GHz accuracy, so 16 channels can share one fiber.

I asked multiple people who build these systems whether you can make DWDM lasers with discrete lasers. The answer was, “Honestly, no.” With discrete lasers, each wavelength comes from a separately made laser whose grating and facet cleave set its wavelength, and a dense grid needs each one held within a tight band around its slot: ±10 GHz or less on a 100 GHz grid.

Where this goes next

The NPO and CPO question will be settled program by program, and so will the number of colors per link. Both come down to whether the parts can be built in volume on production lines that already exist. SHIP™ is our answer to that, and our OCI MSA-compliant part, built to the Optical Compute Interconnect multi-source agreement, is scheduled to exit the fab in the first half of 2027.

If you’re working through an NPO or CPO decision, or a channel plan, I’m glad to compare notes. You can reach us here.