Better Heat Dissipation Points to Improved Electronics
New Materials Simulated on Bridges, Bridges-2 May Allow Smaller Devices, Avoiding “Hotspot” Problem
Progress in electronics has led to smaller and smaller devices. But removing the heat that ever-tinier circuits generate so it doesn’t destroy the components is a looming problem. Scientists used Bridges and Bridges-2 to simulate combinations of materials, preparing them for a successful real-world demonstration of superior heat dissipation in gallium nitride electronics on a boron arsenide substrate compared with current silicon-based materials.
Why It’s Important
In the 1970s, science fiction author Arthur C. Clarke made the then-fantastic prediction that someday each of us would have a high-resolution video screen and a keyboard that let us talk to friends face to face, shop, look up information and any number of activities that linked computer to computer. He was only wrong by not imagining that we could do all of that with a device that fits in the palms of our hands.
Those advances have depended on being able to fit twice as much computing power on a chip about every two years—what’s sometimes called “Moore’s Law.” But Moore’s Law was never really a law. We can’t count on this progress to continue, and one big reason is heat. More-tightly packed electronics generate much higher hotspot temperatures, which can reduce efficiency or even melt the components.
“[Today] there are transistors integrated to a higher density … The power has to dissipate somewhere in the form of heat, or the local temperature in the form of what we call a hotspot is going to increase dramatically. In the 1980s … it was more like a couple hundred degrees Centigrade [around 400°F], like inside an oven. By the 1990s, it’s exceeding that of a rocket nozzle. Today it’s approaching that of the Sun’s surface in some high-power [applications].”—Yongie Hu, UCLA
Yongie Hu and his lab at UCLA have been developing substrates—the slab of material that the electronics sit on—that are better at absorbing and removing heat than current silicon-, silicon-carbide- or diamond-based substrates. To make sure their lab experiments don’t waste time on dead ends, though, they’ve first been simulating the components using PSC’s Bridges and Bridges-2 advanced research computers.
The substrate is a slab of material that an electronic component sits on. Among other things it’s designed to carry heat away from the electronics as fast as possible. The UCLA team used Bridges and Bridges-2 to simulate a new boron arsenate substrate that dissipates heat from new-generation gallium arsenide electronics far better than silicon-based materials or even diamond, which is known for superior heat conductivity.
How PSC Helped
To design better substrates, two numbers were important. The interface thermal boundary conductance, as measured by the Debye temperature for each material, gives a measure of how well the material can transfer heat across connections with other materials. Thermal conductivity is a measure of how well heat travels through a material. More of both is good. It’s also important that these heat-related characteristics of the electronics and the substrate match so that the heat doesn’t get stuck at the connections between the two. These relationships also change as the materials get smaller and act less like bulk materials and more like individual atoms.
The computations that Hu and his colleagues wanted to perform, working from first principles to minimize assumptions, would require massive memory. (That’s the same thing as RAM in a personal computer, only far more of it.) Without big memory, data traffic jams could form within the computer and choke the calculations to death. To get accurate computational results, their advanced theory calculations would use more than a terabyte of computer memory. That’s far beyond a personal computer’s capability, and not possible with most university-run supercomputing “clusters.” Bridges and its successor system, Bridges-2, offered the researchers big memory in the form of 4 to 12 terabytes for each node—and multiple terabyte-scale nodes each. That’s compared with about 16 gigabytes total, 250 to 750 times less, for a good laptop in 2021.
“We’re modeling how the device operates [at] different scales. In this paper in particular, we modeled from 100 nanometers … to 100 microns … So sometimes the supercell [the smallest chunk of material that gets simulated] has a lot of atoms, and [for] some of these existing computational resources … the memory would not be sufficient … Only Bridges would allow us to perform these kinds of calculations … and … gives you very good agreement with experimental [results].”—Yongie Hu, UCLA
Previously, one material that had interested scientists as a promising substrate was diamond. It’s got terrifically high Debye temperature and thermal conductivity—better than the silicon carbide currently used in high-density electronics. But diamond’s Debye temperature isn’t that close to those for silicon circuitry or newer, better-heat-conductive candidate electronics like gallium nitride (GaN). To get a better match with the electronics, Hu’s team has been investigating boron arsenide (BA) and boron phosphide as substrates. The new substrates have over three times higher thermal conductivity than silicon carbide substrates. They also have a much better thermal boundary conductance match with gallium nitride electronics.
The simulations on Bridges and Bridges-2 suggested that both boron substrates would be a good match for gallium nitride electronics at multiple size scales. The predictions also had excellent agreement with results when the materials were tested in the lab. When compared with substrates made of diamond or silicon carbide, boron arsenide draws heat from gallium nitride much more effectively. Though diamond was a little better than silicon carbide, boron arsenide was far better, reaching much lower temperatures for all running power densities. For example, the GaN-BA combination reached 60 °C, compared with 110 °C for GaN-diamond or 140 °C for GaN-SiC at less than 15 watts per millimeter of circuit.
The scientists reported their results in the journal Nature Electronics in June 2021. They’re continuing to develop these materials for industrial use as components in even smaller electronics and radio-frequency circuits. A February 2021 Nature Communications report from the team successfully demonstrated self-assembling of the material as high-performance flexible thermal interfaces for wearable electronics and soft robotics.