xMEMS Introduces Solid-State Micro-Cooling to Edge AI and Data Centers

As compute proliferates across everyday devices, cooling remains a major design challenge. To solve that, xMEMS Labs today announced the start of mass production for its XMC-2400 and XMC-4800 µCooling chips. Both solid-state air movers build on the company’s PiezoMEMS platform and aim to bring active cooling to on-device AI platforms like glasses and smartphones.

Rendering of xMEMS cooling a smartphone. Image used courtesy of xMEMS

All About Circuits spoke with Mike Housholder, VP and GM of thermal management at xMEMS, to learn how silicon air movers could give AI hardware room to keep scaling.

The New XMC-2400 and XMC-4800

Thin, densely packed devices like AI glasses and smartphones leave designers little space for a rotary fan and even less tolerance for its noise. Conventional fans also depend on magnets, copper coils, and spinning blades, all of which add weight and thickness. To work around those limits, xMEMS designed the XMC-2400 and XMC-4800 around an array of PiezoMEMS cells on a monolithic die. Each cell acts as an air pulse generator. Housholder described the result simply.

“It's an air mover on a chip,” he explained. “By arraying these cells, we can generate a positive airflow and basically use silicon to move air and create cooling.” 

Because the MEMS actuators operate in the ultrasonic range, the company says the chips produce no audible motor noise.

Rendering of airflow direction using PiezoMEMS. Image used courtesy of xMEMS

Building on that architecture, xMEMS sized the two parts for different thermal loads. The XMC-2400 measures 7.42 mm x 9.48 mm x 1.13 mm and draws approximately 150 mW. It delivers up to 28 cc/s of airflow at up to 1,300 Pa of back pressure. For heavier loads, the XMC-4800 measures 9.93 mm x 14.35 mm x 1.13 mm and draws approximately 240 mW. In exchange, the larger part moves up to 48 cc/s at up to 1,100 Pa. Those back-pressure figures seem a deliberate design choice.

“This device is more of a pump versus a fan," Housholder said. “We can overcome a lot of resistance, and we can move air in very compact, tiny spaces."

According to xMEMS, customers have measured chip and system-level temperature reductions of 10°C to 25°C in glasses, smartphones, and SSD designs. Meanwhile, to simplify mechanical integration, xMEMS routes airflow vertically through the package and lets designers reverse its direction on the fly. Air passes through slits in the metal shield on top and through holes in the LGA substrate below. For layouts that need a 90° turn, the company also offers side-vented "-S" variants. Engineers can then reflow the parts onto rigid or flex PCBs with standard pick-and-place equipment.

Because the chips eject hot air, designers still need an inlet and an outlet in the enclosure. To protect other components, xMEMS recommends a dedicated duct that couples thermally to the hot component while staying physically isolated. The company rates the parts IP58 at the component level and says a protective mesh lets OEMs reach IP68 at the system level.

Why Back Pressure Matters in Dense Enclosures

Thermal designers generally characterize fans with a P-Q curve, which plots static pressure against volumetric flow rate. They then overlay the system impedance curve, and the intersection marks the fan's actual operating point.

As component density rises, the impedance curve steepens and pushes that operating point toward lower flow. In one study of installed axial fans, researchers found that maximum flow drops once the enclosure inlet area falls below twice the fan flow area. Phone and glasses designers rarely have that much opening to work with. In those conditions, a conventional fan can stall well short of its rated airflow.

A fan's operating point sits where its P-Q curve meets the system impedance curve. Image used courtesy of Journal of Environment
and Engineering

Without forced airflow, engineers fall back on graphite sheets and vapor chambers to spread heat through the chassis. Spreading lowers hotspot temperatures, but the heat still leaves through the device surface. In skin-contact products, OEMs typically cap surface temperatures in the low-to-mid 40°C range and throttle processors to stay under that limit. Once spreaders saturate the available surface area, designers need some way to move hot air out of the device.

Targeting Thermal Orphans in the Data Center

Similar constraints appear inside servers, where liquid cooling handles the largest heat sources but leaves smaller ones behind. Operators deploy cold plates on CPUs and GPUs, yet 5- to 10-W components on the far side of the board may not justify their own liquid loop.

“There will be some thermal orphans, where it's just not feasible or fiscally responsible to route liquid cooling to every component,” Housholder said.

For E1.S and E3.S enterprise SSDs, designers may mount a cold plate on one side while NAND sits on both sides of the PCB. The NAND under the cold plate runs cooler than the NAND opposite it, and that gradient caps drive throughput.

“That warmer NAND is going to dictate your data rates," Housholder said. 

To shrink that gradient, xMEMS says its first data center SSD customer is placing µCooling on the drive's non-cold-plate side.

From Qualification to Volume Shipments

With mass production underway, xMEMS expects the first commercial XMC-2400 products in early 2027. Beyond SSDs, xMEMS is eyeing optical transceivers as co-packaged optics, pulling high-power elements closer to the GPU.

"What's left is the laser module, because that's a high-maintenance item,” Housholder said. “That, for now, is remaining as a pluggable, and the power density there is right in our wheelhouse."