In recent years, data center rack density has increased rapidly. To meet the demands of artificial intelligence and other high-performance computing applications, operators are cramming more computing power into each server rack. This means each rack needs to consume more kilowatt-hours of electricity, ultimately generating more heat. Cooling infrastructure is struggling to keep up with this trend.

Microfluidics Technology

David Holmes, Chief Technology Officer for Global Industry at Dell Technologies, stated, "Eight years ago, the average power of a rack was 6 kilowatts, while racks now ship at 270 kilowatts. Next year, 480-kilowatt racks will be available, and megawatt-class racks will be on the market within two years." Swiss company Corintis is developing a technology called microfluidics, which delivers water or other coolants directly to specific areas of a chip to prevent overheating. In a recent test conducted in collaboration with Microsoft, servers running Microsoft Teams video conferencing software cooled the chip three times more efficiently than other existing cooling methods. Compared to traditional air cooling, microfluidics can reduce chip temperatures by more than 80%.

Improving chip performance using microfluidic technology

Lower chip temperatures allow chips to execute instructions faster, improving performance. Chips operating at lower temperatures are also more energy-efficient and have lower failure rates. Furthermore, the temperature of cooling air can be increased, reducing the need for coolers, lowering liquid consumption, and thus improving the energy efficiency of data centers.

By precisely delivering coolant to the hottest areas of the chip, the amount of water required for chip cooling can be significantly reduced. Van Erp points out that the current industry standard is approximately 1.5 liters of water per kilowatt of power per minute. As chip power approaches 10 kilowatts, this means that cooling a single chip will soon require 15 liters of water per minute - a figure that will undoubtedly anger communities concerned about the environmental impact of the massive "AI factories" (potentially containing millions or more GPUs) being built in their areas.

"We need liquid cooling optimized for chips to ensure that every drop of liquid reaches the right place," says Remco van Erp, co-founder and CEO of Corintis. Corintis develops simulation and optimization software used to design microchannel networks on cold plates. Just like the arteries, veins, and capillaries in the human circulatory system, the ideal cold plate design for each chip is a complex network of precisely shaped channels.

Corintis has expanded its additive manufacturing capabilities to mass-produce copper components with channels as thin as a human hair (approximately 70 micrometers). Its cold plate technology is compatible with today's liquid cooling systems.

The company believes this approach can improve the heat dissipation of cold plates by at least 25%. Corintis believes that by working directly with chip manufacturers to etch channels onto silicon wafers, a tenfold improvement in heat dissipation performance can ultimately be achieved.

Advancing the development of liquid cooling technology for artificial intelligence chips

Liquid cooling technology is not new. For example, the IBM 360 mainframe used water cooling more than half a century ago. Modern liquid cooling technologies are mainly divided into two types: immersion systems (immersing the rack or even an entire row of devices in coolant) and direct chip cooling systems (delivering coolant to a cold plate close to the chip).

Immersion cooling technology is still in its early stages. While direct chip cooling technology has been widely used for GPU cooling, it only cools the chip surface.

Van Elp stated, "Current liquid cooling technology is a one-size-fits-all solution, relying on a simple approach that is not optimized for the chip, which hinders good heat transfer. The optimal design for each chip is a complex network of precisely shaped microchannels that match the chip and guide the coolant to the most critical areas." Corintis has partnered with chip manufacturers to improve chip design. Chip manufacturers use the company’s thermal simulation platform to program the heat dissipation on silicon test chips at millimeter-level resolution and test the final temperature of the chip after installing the selected cooling scheme. In other words, Corintis acts as a bridge between chip design and cooling system design, enabling chip designers to build chips for future AI applications with superior thermal performance.

The next goal is to transition from bridging cooling channels and chip design to integrating these two processes into one. "Modern chips and cooling are currently two separate parts, and the interface between them is one of the main bottlenecks in heat transfer," said Van Elp.

To improve cooling performance tenfold, Corintis is betting on the future trend of tightly integrating cooling with the chip itself - microfluidic cooling channels will be etched directly inside the microprocessor package, rather than onto an external cold plate.

Corintis has already produced over 10,000 copper cold plates and is ramping up production, aiming for an annual output of 1 million units by the end of 2026. The company has also developed a prototype production line in Switzerland to develop cooling channels directly inside the chip, rather than on a cold plate. This technology is currently planned for small-batch production to validate the basic concept before being rolled out to chip manufacturers and cold plate producers.

Corintis announced these expansion plans immediately after Microsoft released the Teams beta. In addition, the company will open an office in the United States to serve American clients and establish an engineering office in Munich, Germany. The company also announced the completion of a $24 million Series A funding round led by BlueYard Capital and other investors.

Source: Content compiled from IEEE

View more at EASELINK