OCP Global Summit 2026 | Ayar Labs, Wiwynn | Nandita Aggarwal, Willie Chen | Rack-Ready High-Power ELSFPs

1. Why the Laser Module Has to Be Designed With the Cold Plate

This short talk at the OCP Future Technologies Symposium ran about 12 minutes, and Aggarwal opened by naming the worry shared by the speakers before her: optics is harder than copper, and the hard part is serviceability.

Ayar Labs splits the architecture in two: the optical engines sit in multi-chip packages (MCP) at the rear of the chassis, right next to the ASICs and GPUs; the laser is built as an ELSFP (external laser source small form-factor pluggable) on the front panel, sending continuous-wave light over fiber to the engines in the back. The reasoning is simple: the laser is the weakest and most heat-sensitive element in the link, so it goes where it is easiest to replace.

But once the laser leaves the package for the front panel, the system decides its thermal environment. The split of work is on the slide: Wiwynn owns the L11 liquid-cooled rack, the L10 compute tray and the cold plate, and first set the boundary conditions (coolant flow, temperature, pressure, splits in the cooling architecture, thermal resistance stack-up); Ayar Labs then designed the laser module to those conditions.

▲ From Wiwynn's L11 liquid-cooled rack, L10 compute tray and cold plate all the way to Ayar Labs' 8 × 23 dBm ELSFP and optical engine (Image source: Ayar Labs, Wiwynn | 2026 OCP Global Summit presentation (p. 2))

Rack-level CPO architecture is not the focus here. What matters is the last box in the chain: how to make the thermal interface reliable for a laser module that sits on the front panel and gets plugged and unplugged repeatedly.

2. Breaking Down the Optical Data: Where the Margin Is, and Where It Isn't

The test conditions were deliberately harsh: 50°C laser die, every laser on (full thermal load), 45°C coolant inlet, and -20 dB optical back-reflection injected at the system side to emulate reflections such as a fiber break. The distributions across 100+ modules:

Parameter	Mean	Std. dev.	Worst case	Spec	Margin (STT calc.)
Output power (in fiber)	23.4 dBm	0.32 dB	Lowest ~22.2 dBm (STT chart read)	23 dBm target	Mean +0.4 dB
SMSR	53 dB	1.5 dB	> 47 dB	> 40 dB	At least 7 dB
Linewidth	0.16 MHz	0.04 MHz	< 0.37 MHz	< 1 MHz	About 2.7x
Average RIN	-158 dB/Hz	0.8 dB/Hz	< -154 dB/Hz	< -150 dB/Hz	At least 4 dB

▲ Distributions of power, SMSR, linewidth and RIN for 100+ ELSFPs at a 50°C die temperature and -20 dB back-reflection (Image source: Ayar Labs | 2026 OCP Global Summit presentation (p. 5))

The slide's line "every laser meets spec with margin" lists SMSR, linewidth and RIN, but not power. The power histogram has a small tail to the left of the 23 dBm target line, and the spoken version was that "most" ELSFPs exceeded the power target. Using a normal approximation with a 23.4 dBm mean and 0.32 dB standard deviation, roughly one laser in ten (about 11%) falls below 23 dBm (STT estimate). Measured as "how many standard deviations the mean sits from spec," linewidth is at about 21, RIN about 10, SMSR about 8.7, and power only 1.25 (STT calculation).

▲ SMSR, linewidth and RIN all sit well clear of spec, while output power averages only 0.4 dB above the 23 dBm target (Image source: Simple Tech Trend original, data from Ayar Labs 2026 OCP Global Summit presentation)

This is not a failure; it is a signal of where the pressure sits: spectral quality is no longer the ELSFP challenge, power is. In Q&A the speaker confirmed that 23 dBm is power "in the fiber, after the host connector," which works out to about 200 mW per channel and about 1.6 W across 8 channels per module; at the 23.4 dBm mean (about 219 mW) it is about 1.75 W of optical power (STT calculation). We discussed the trade-off between power and split count in CPO lasers don't have to chase 400 mW: NPO integrated lasers; this data shows that delivering 200 mW in fiber reliably at 50°C with back-reflection still takes screening.

3. Breaking Down the Thermal Interface: 45°C Coolant, 50°C Die

The two most important numbers in the talk are temperatures: 45°C coolant, 50°C laser die. 45°C is the rack's design inlet temperature; the laser must hold wavelength and power at 50°C, leaving just 5°C of room, with the active TEC inside the module handling the rest.

The TEC can lock each laser at its set point, but it cannot control the temperature of the module's outer surface. If that surface is uneven, the TEC works harder; Aggarwal's phrase was keeping the system in "a predictable picojoule-per-bit environment." The two companies ran joint CFD across the cold plate and the module and solved two challenges:

• Limited contact area: the ELSFP sits inside a cage with only a top opening for cold-plate contact, and a soft TIM gets scraped during insertion. The fix is a graphite resilient TIM, protected by a beryllium-copper spring sheet underneath and pressed against the copper plate. In Q&A the design pressure was given as about 50 psi and insertion force about 14 N, with the TIM a graphite and silicone-based compound pad that keeps some rebound (spoken figures, not on the slides).

• Heat build-up at the center: with uniform fin pitch, the middle modules ran hotter, giving a 2.9°C surface variation. Localized fin pitch redistributed the flow and cut the variation to 0.8°C, about 3.6x (STT calculation).

▲ Two design challenges: thermal contact through the cage opening, and heat buildup at the center of the cold plate; a graphite compliant TIM with beryllium-copper springs and localized fin pitch cut the temperature spread from 2.9°C to 0.8°C (Image source: Wiwynn, Ayar Labs | 2026 OCP Global Summit presentation (p. 7))

Validation: 20 ELSFPs, 45°C inlet, 1.3 LPM flow, with laser die temperature within 50 ± 1°C on all of them (the slide chart shows only the 10 top-side modules; the bottom side is said to follow the same pattern).

▲ The thermal interface stack from 45°C coolant to the 50°C laser die: copper cold plate, localized fins, graphite compliant TIM, beryllium-copper springs, cage opening and TEC (Image source: Simple Tech Trend original, data from Ayar Labs, Wiwynn 2026 OCP Global Summit presentation)

Put simply: the laser vendor owns the optical specs, but whether the laser holds 50°C is half the cold-plate vendor's design. That is why the two companies presented together.

4. Twenty Modules in 1OU: The Price of Serviceability

Wiwynn's cold plate uses top and bottom plates in a belly-to-belly layout, cooling 20 ELSFPs within 1OU height, each side with its own inlet and outlet. That works out to 160 laser channels per 1OU, about 35 W of in-fiber optical power at the 23.4 dBm mean, or about 32 W at the 23 dBm target (STT calculation).

The key to serviceability is the TIM: hundreds of repeated insertions with no wear-out, and no re-application in the field. Conventional thermal paste or soft pads do not survive repeated plugging on a pluggable module; without this, "front-panel replaceable" is replaceable only in form factor. We looked at switch-side thermal trade-offs in 51.2T NPO switch thermal design; this talk shrinks the same problem down to a single laser module.

5. Risks and Counterarguments

• Tight power distribution: the mean is only 0.4 dB above target, and roughly one laser in ten may be below 23 dBm (STT estimate). Whether volume production relies on screening, binning or higher drive current was not said.

• Module-to-module power differences: on the validation chart for the 10 top-side modules, "ELSFP Power Delta" is about 0–6% for #1–9 but about 19–20% for #10 (STT chart read); neither the metric definition nor the cause was explained.

• Unclear surface-temperature measurement point: the same chart shows ELSFP surface temperature around 40–41°C (STT chart read), below the 45°C coolant supply; along a passive cold-plate heat path the module housing should not run colder than the coolant, so the measurement point or legend definition needs checking against official material.

• Cost not broken out: asked about the incremental cost of the ELSFP subsystem, Ayar Labs said it is not separated from its optical engine cost model.

• The interface is still a bilateral agreement: parameters like 50 psi and 14 N were only mentioned verbally in Q&A (per auto-captions); whether a different cold plate or module vendor would use the same numbers has no spec to rely on today.

6. What It Means for Taiwan's Supply Chain

• Wiwynn gets definition rights, not just assembly. The slide states that Wiwynn set the boundary conditions first and Ayar Labs designed the laser to them. In a CPO architecture, the ODM moves from "putting parts into the rack" to "specifying the thermal environment the laser module must work in."

• Thermal parts extend from fans and cold plates to the optical module interface. Graphite compound TIMs, beryllium-copper spring sheets and localized-fin cold plates are processes familiar to Taiwan's thermal and mechanical supply chain, but no public information shows who supplies them, so no supplier should be assumed.

• The laser die is multi-sourced by design. Ayar Labs' slide states the ELSFP uses "multi-vendor laser die," which opens a door at the die level; the bar is a tight in-fiber 23 dBm distribution at 50°C with -20 dB back-reflection. For spec context, see Meta, Broadcom and AMD define the OCI 200G line-side spec.

7. Conclusion

The talk set no new dBm record; its value is in reframing the ELSFP question: not whether the laser is strong enough, but whether every laser plugged into the front panel, inserted and removed repeatedly and sitting in a 45°C liquid-cooled environment, can hold 50°C. Ayar Labs and Wiwynn answered yes with 100+ modules and a 20-module full-load validation, but the answer is tied to a bilateral thermal interface.

Three things to watch over the next 12–18 months: whether OCP opens a work item on ELSFP thermal interface and cold-plate requirements, whether the 23 dBm power distribution shifts up in production lots, and whether a second cold-plate or module vendor can reuse the same interface parameters.

Moving the laser from the package to the front panel buys serviceability; the price is that the thermal interface turns from a component problem into a system spec.

This article is for technology and industry trend analysis only and does not constitute investment advice.