In data centers, pluggable optical transceivers convert electronic bits to photons, fling them across the room, and then turn them back to electronic signals, making them a technological linchpin to controlling the blizzard of data used in AI. But the technology consumes quite a bit of power. In a data center containing 400,000 GPUs, Nvidia estimates that optical transceivers burn 40 megawatts. Right now, the only way to deal with all that heat is to hope you can thermally connectthese transceivers to the switch system’s case and cool that. It’s not a great solution, says Thomas Tarter, principle thermal engineer at startup xMEMs Labs, but because these transceivers are about the size of an overlarge USB stick, there’s no way to stick a conventional cooling fan in each.
Now, xMEMs says it has adapted its upcoming ultrasonic microelectromechanical (MEMS) “fan-on-a-chip” to fit inside a pluggable optical transceiver so it drives air through and cools the transceiver’s main digital part, the digital signal processor (DSP). Keeping the DSP cool is critical to its longevity, says Tarter. At upwards of US $2,000 per transceiver, getting an extra year or two from a transceiver is well worth it. Cooling should also improve integrity of the transceivers’ signals. Unreliable links are blamed for extending already-lengthy training runs for new large language models.
xMEMS’ Cooling Tech Finds a New Home
The xMEMS chip cooling tech, which was unveiled in August 2024, builds on the company’s earlier product, solid-state microspeakers for earbuds. It uses piezoelectric materials that can change shape at ultrasound frequencies to pump 39 cubic centimeters of air per second through a chip just about a millimeter high and less than a centimeter on a side.
Smartphones, which are too slim to carry a fan, were the first obvious application for the MEMScooler, but cooling the fast-growing data-center-scale AI systems seemed out of reach for MEMS technology, because it can’t come near matching the liquid cooling systems removing thousands of watts of heat from GPU servers.
“We were pleasantly surprised by the approach by data center customers,” says Mike Housholder, xMEMS vice president of marketing. “We were focused on low power. So we didn’t think we had a slam dunk.”
Pluggable optical transceivers turn out to be a data center technology that is squarely in the fan-on-a-chip’s wheelhouse. Today, heat from a transceiver’s DSP, photonics IC, and lasers is thermally coupled to the network switch computers they are plugged into. (These usually sit at the top of a rack of computers.) Then air moving over fins built into the switch’s face remove the heat.
In collaboration with partners they would not name, xMEMS began exploring how to get air flowing through the transceiver. These parts consume 18 watts or more. But by situating the company’s MEMS chip within an airflow channel that is thermally connected to the transceiver chips but physically isolated from them, the company predicts it will be able to drop the DSP’s temperature by more than 15 percent.
xMEMS has been making prototype MEMS chips at Stanford’s nanofabrication facility, but it will have its first production silicon from TSMC in June, says Housholder. The company expects to be in full production in the first quarter of 2026. “That aligns well with our early customers,” he says.
Transceiver shipments are growing fast, according to Dell’Oro Group. The market analyst predicts that shipments of 800 gigabit per second and 1.6 terabit per second parts will grow at more than 35 percent per year through 2028. Other innovations in optical communications that could affect heat and power are also in the offing. In March, Broadcom unveiled a new DSP that could lead to a more than 20 percent power reduction for 1.6 Tbps transceivers, due in part to use of a more-advanced chip manufacturing process. The latter company and Nvidia, separately, have developed network switches that do away with pluggable transceivers altogether. These new “co-packaged optics” do the optical/electronic conversion on silicon within the switch chip’s package.
But Tarter, who has been working on cooling chips since the 1980s, predicts there will be more applications both inside and outside the data center for the MEMS chip to come. “We’re learning a lot about applications,” he says. “I’ve come up with 20 or 30 basic applications for it, and hopefully that inspires designers to say ‘Oh, this is how I can use this in my system.’”
From Your Site Articles
Related Articles Around the Web
