Mechanical engineers have designed a more effective and energy-efficient technology for cooling computer chips. Publishing May 7 in the Cell Press journal Cell Reports Physical Science, the researchers used a mathematical algorithm and advanced 3D printing method to produce pure copper cold plates that outperformed conventional cold plates and required less energy to run. If used to cool an entire data center, the technology would contribute only around 1.1% of the data center's total energy usage compared to more than 30% for conventional air-cooling methods, the researchers estimate.
"Cooling is the bottleneck in computer-chip design," says first author Behnood Bazmi, a mechanical engineer at the University of Illinois Urbana-Champaign, USA. "By bridging the gap between computational design and manufacturing capability, our approach provides a pathway for more energy-efficient liquid cooling of chips and other electronics."
Computer chips are becoming increasingly high powered, which means they produce more heat. This, combined with the increase in data centers, is putting a strain on the energy grid—by 2028, it's predicted that data centers will consume up to 12% of the national grid load in the United States. For the past 40–50 years, computer chips have been cooled by circulating air, but air is insufficient for dissipating the heat produced by modern chips. Liquid direct-to-chip cooling could offer a more effective solution, say the researchers.
Direct-to-chip cooling systems consist of a cold plate that is attached to a computer chip. These cold plates have tightly packed metal "fins" that project out into the cooling liquid to maximize the surface area that is in contact with the coolant. Some direct-to-chip liquid cooling systems are already commercially available, but those systems prioritize manufacturing cost over performance. In this study, the researchers set out to optimize fin design to design cold plates with maximal cooling ability.
The team used a technique called topology optimization to design fins with an optimal shape. From a simple rectangular starting design, topology optimization uses a mathematical algorithm to gradually alter the fin's shape. For each iteration in fin design, the algorithm estimates the cooling capability and the amount of power that would be needed to pump coolant past the fins.
"Topology optimization ends up converging on a design which is optimal in maximizing thermal performance and minimizing pumping power," says senior author and mechanical engineer Nenad Miljkovic of the University of Illinois Urbana-Champaign.
With pointed tops and jagged edges, the resulting fins are much more complex than conventional fins, which are usually simple rectangles, cones, or cylinders. Because this design would be too difficult to manufacture using conventional techniques, the team collaborated with a company called Fabric8 to use an advanced manufacturing method called electrochemical additive manufacturing (ECAM) to produce copper cold plates with the optimized fins. Rather than melting copper, ECAM relies on electrochemical plating to deposit copper and build the fins up, layer by layer, from bottom to top.
Pure copper has a high thermal conductivity, but it's difficult to 3D print, so most cold plates are made of an aluminum alloy (AlSiMg) or stainless steel, which are not optimal for heat transfer. "ECAM can manufacture pure copper parts with very fine detail—down to 30 to 50 micrometers, less than the width of a human hair," says Miljkovic.
When the researchers compared the cooling performance of an individual copper cold plate with the optimized fins to cold plates with conventional rectangular fins, they found that the optimized plate delivered up to 32% better cooling and reduced pressure drop (less effort to push fluid through the cold plate) by up to 68% while maintaining the same cooling performance. At the level of an entire data center, this would translate into significant energy savings compared to both air-cooling and commercially available liquid-cooling systems, the researchers say.
For example, a data center with 1 gigawatt (GW) of computing power consumes around 550 megawatts to run an air-cooling system, meaning it actually consumes 1.55 GW total in energy, but only 1 GW is used for functions such as ChatGPT, searches, and storage. "With our cold plates, data centers would only need to use 11 megawatts for cooling instead of 550 megawatts," says Miljkovic.
This optimization and manufacturing system could be scaled to design optimized cooling systems for other electronics and non-electronic applications, the researchers say. "Our workflow can be applied to a wide range of cooling challenges across different length scales," says Bazmi.
