Microsoft is making advances with a new way to cool microchips that it says could lead to more energy-efficient data centers in the future. It’s a method called microfluidics that involves liquid coolant flowing directly into the silicon.
After lab tests, Microsoft found that this strategy can remove heat up to three times better than cold plates currently used in data centers today. The company announced this week that it was able to develop a microfluidic cooling system for a server running core services for a simulated Microsoft Teams meeting.
If they can find the same success outside of a lab, microfluidics could cut down the amount of energy needed to cool a data center. It could also lead to more powerful chips that current cooling systems would struggle to keep from overheating. But there are still a lot of factors that could influence how impactful this new technology ultimately is in the real world.
It could lead to more powerful chips that current cooling systems would struggle to keep from overheating
Compared to data centers of yore, the next generation being built to train and run new AI models house more powerful chips. Not only do those GPUs use a lot of energy, they also get very hot. Keeping them cool is a challenge that not only affects performance, but also makes data centers consume more energy.
Typically, a data center might use fans to pass cool air over a chip. A more advanced technology that Microsoft employs for higher-powered chips involves cold plates made of copper with fluid flowing through them. Put that plate on top of a chip, and it whisks away the heat.
With microfluidic cooling, liquid flows through channels etched onto the back of a chip. The trick is making sure the channels, about the width of a human hair, are deep enough to prevent clogging but not so deep that the chip becomes more likely to break. Microsoft says it used AI to figure out where to direct coolant onto a chip to chill it most efficiently. The etched designs are also inspired by nature — mimicking the patterns of veins on leaves, for example — that have already shown how practical they are at distributing water and resources. Using microfluidics, Microsoft documented a 65 percent reduction in the maximum temperature rise of the silicon of a GPU.
The advantage with microfluidics is that it brings fluid straight to the chip, eliminating the need for protective layers of materials between the chip and the coolant when cold plates are used. Each layer, like a blanket, holds in some heat, and so the coolant needs to stay colder to work well within cold plates. Cold liquid flows into the plate; hot fluid flows out and needs to be cooled down again. With microfluidics, the coolant doesn’t need to be chilled to as low of a temperature, conserving energy.
Microfluidics can also allow a data center to more efficiently handle peaks in demand. Teams calls usually start every hour or half-hour, Microsoft gives as an example. To handle those spikes in demand, they might have to install more servers to have enough capacity on hand even if they won’t be used all the time. The alternative would be to let existing servers work extra hard, called overclocking — but that could lead to overheating and damaging the chip. Microfluidic cooling, because it’s more efficient, can allow for more overclocking without the same risk of a chip melting down.
In theory, if servers can work harder than they do now without melting chips down, a data center might not need as many of them. And by minimizing the risk of overheating, microfluidics could also allow for more tightly packed servers within a single data center. It could cut the costs, literal and environmental, of needing to build additional facilities.
All these benefits could be key for next-generation microchips, which are expected to become so powerful that cold plates may fall short. Microsoft says that microfluidics could also enable 3D chip architecture. 3D chips would be even more powerful than today’s semi-flat designs, but heat has been a stumbling block for making this happen. With microfluidics, however, there’s the possibility of flowing coolant through the chip.
Efficiency can also be a double-edged sword
Microsoft doesn’t have a timeline for when this all might happen. After more lab testing comes the challenge of figuring out how to make the hardware and supply chain changes needed to allow for microfluidics — for example, at what point in the manufacturing process will grooves be etched into the chips? Fortunately, they can use the same kind of coolant, a mix of water and propylene glycol, used today in cold plates.
Other researchers have been studying microfluidics for years also. HP, for example, was awarded $3.25 million in funding from the Department of Energy last year to develop its own microfluidic cooling technology. “All these things are good to see, we’re happy to see them, and where we can participate to move things faster we’re happy to,” says Husam Alissa, director of systems technology in Cloud Operations and Innovation at Microsoft.
Microsoft says it “hopes to help pave the way for more efficient and sustainable next-generation chips across the industry” in its recent blog post touting its progress on microfluidics. Energy efficiency is crucial if the company wants to operate more sustainably. Like other tech companies, Microsoft’s planet-heating carbon emissions have grown as it leaned into generative AI. But efficiency can also be a double-edged sword. As something becomes more efficient and affordable to use, people tend to use much more of it and that could ultimately lead to an even bigger environmental footprint. It’s a phenomenon called the Jevons paradox, which even Microsoft CEO Satya Nadella has commented on as a force driving greater adoption of AI.
