Vaire, a U.K. startup that is working on a computer chip technology that could radically reduce the amount of energy required to run AI workloads, says initial tests of its new chip components show that it could halve the amount of electricity required to run many computations, including those used in artificial intelligence.
If the results hold up—and they have yet to be independently verified, although test kits of Vaire’s chip are currently en route to potential customers and academic labs—the breakthrough could lead to significant commercial adoption of the unusual technology, which is known as reversible computing. Ultimately reversible computing holds out the prospect of chips that could recycle almost all the energy they require and emit almost no heat.
The technology behind Vaire’s chips could address one of the most significant challenges associated with the AI boom: the amount of energy AI requires. Electricity demand from data centers full of graphics processing units, the specialized computer chips typically used to train and run AI models, is soaring. Microsoft has even cut a deal with Constellation Energy to restart a mothballed nuclear reactor at the notorious Three Mile Island nuclear power plant in Pennsylvania entirely to supply one of the software firm’s new AI data centers.
Globally, the International Energy Agency has forecast that electricity demand from datacenters will more than double by 2030 to 945 terawatt-hours, or about 3% of global energy demand and an amount equivalent to the energy that all of Japan currently consumes. AI workloads will be the biggest driver of that growth, with the energy demands for AI workloads in data centers set to quadruple in the next five years. What’s more, many AI industry insiders think the IEA’s projections are too conservative.
That energy consumption is also raising substantial fears about the impact AI will have on the environment. The carbon footprint of AI systems is difficult to estimate, since it depends largely on the power sources for any particular data center. Many AI data centers in the U.S. and Europe are being built using renewable energy, such as solar and wind power, or nuclear power, which does not emit much carbon dioxide, but in other regions, substantial amounts of gas or coal-fired power are being used, increasing CO2 emissions. There is also a concern that tech companies are purchasing so much green power for their data centers, that it is forcing other potential customers to fall back on fossil fuels.
The amount of water used to cool AI data centers is also an increasing concern. Because GPUs draw more power than other kinds of computer chips, they also generate far more heat, meaning it takes more energy and more water to cool data centers packed with GPUs. And many AI data centers are being built in areas where ground water supplies are already stretched.
Three years operating in stealth
It is this challenge serial entrepreneur Rodolfo Rosini set out to solve when he started Vaire in 2021. He said it was obvious even then that AI models were getting larger and larger, requiring ever greater amounts of energy and that AI was going to be “very big, very soon,” he says.
To Rosini, the obvious solution was reversible computing. (More on what that is exactly in a moment.) He co-founded Vaire with Hannah Earley, a University of Cambridge researcher who specializes in unconventional computing and Andrew Sloss, a former principal researcher at the chip design company Arm.
For three years, Vaire operated in stealth, with just $500,000 in seed funding from London-based 7percent Ventures and a small group of angel investors. It used that time to perfect the design of its reversible computing chip. Last year, the company raised additional seed funding from 7percent Ventures, Lifeline Ventures, SeedCamp and Jude Gomila, the co-founder of mobile ad company Heyzap (which was acquired by Fyber). The company has raised $10m total, a number not previously reported. The company currently employs about 20 people in offices in London, Cambridge, England, Sunnyvale, California, and Portland, Oregon.
The graphics processing units (GPUs) used for AI produce so much heat because they draw more electricity than other kinds of computer chips, such as the central processing units (CPUs) that power traditional servers. But almost 100% of the electricity that any microprocessor chip (whether a GPU or CPU) uses to perform computations is converted to heat and lost (an old joke among engineers says a computer is simply a toaster that happens to perform calculations). Most of that heat is generated when the circuit overwrites the information it is currently holding in its circuits with the next set of information it needs to calculate.
A technology decades in the making
Physicists have known since at least the early 1960s that if the logical process in a computer circuit could be reversed—so that the information in the circuit is restored to its original state rather than being simply overwritten—it would be possible to conserve and recycle the vast majority of the electricity a chip uses. The problem was that it was an extremely difficult engineering challenge to create a chip that could do this, especially using the same silicon-based CMOS (complementary metal-oxide semiconductor) fabrication process that is used to build all modern computer chips. Nonetheless, researchers at MIT overcame many of these obstacles and built working prototype reversible computing chips using CMOS in the late 1990s.
At the time though, there was little impetus to implement reversible computing commercially. Moore’s Law—which stipulated that the number of transistors that can be packed on a given area of silicon was doubling every two years—showed no signs of ending. And Moore’s Law had an important corollary known as Koomey’s Corollary: as the size of transistors kept shrinking, the amount of electricity needed to perform any given computation halved every 1.6 years (a rate even faster than Moore’s Law itself.) As long as those trends kept up, there was little need to be worried about computer chips’ hunger for power. “Increased energy efficiency was coming for free along with the increased density of transistors on the chip,” Neal Anderson, a professor of computer and electrical engineering at the University of Massachusetts Amherst, said.
But now Moore’s Law has run into a number of physical limitations. The most advanced chip fabrication processes can print transistors at a pixel resolution as small as two nanometers, or one-forty thousandth the width of a human hair. At that size, transistors require so little energy to store one bit of information that there are significant risks of quantum mechanical effects that introduce random errors into the chip’s calculations. Making the transistors any smaller means those effects will render them unreliable.
Heat is another problem. Chip designers have kept the clock speeds—measured in GHz—at which the transistors on a chip cycle through states more or less constant since 2005. That’s because cycling the circuits any faster produces so much heat it risks damaging the transistors—literally frying the circuits. The heat can also interfere with the transmission of the signal through the circuit, introducing errors. What’s more, at higher temperatures, electrons actually move more slowly, limiting how fast the circuit can operate. To increase processing power at constant cycle speeds, chip designers have instead added many more processing cores that can handle different parts of a computation in parallel, rather than trying to have each core perform more computations per second. This has allowed engineers to continue to boost overall chip performance, but it has made it more difficult to make the chips more energy efficient.
Reversible computing offers a solution to this. “Reversible computing is a scientifically sound way going forward,” Brian Tierney, a researcher in advanced computing at Sandia National Laboratories in Albuquerque, New Mexico, said.
An engineering challenge
Rosini said when he first thought of setting up Vaire, he approached many of the world’s leading chip companies, assuming they already had teams working on reversible computing. He said, to his surprise, none of them were actively pursuing it. Only one company, he said, seemed to have even considered it, but told Rosini its research teams had too many other priorities to devote much energy and money to trying to make it work. “They said it was too far down the list,” he said.
Engineering a reversible computing chip, especially using conventional chip fabrication methods, is tricky. Rosini eventually recruited Mike Frank, a researcher who has spent most of his career working on reversible chip designs, first at MIT and the University of Florida and later at Sandia National Laboratories, to help with Vaire’s design.
To recycle the electricity back through the circuit requires marrying an analog component known as a resonator with the digital ones normally found on a chip. Radios have long used components such as resonators, but the ones Vaire needed to create for reversible computing have to generate an unusual signal shape—a trapezoid rather than a traditional sine wave. Perfecting such a resonator was one of the key engineering challenges Vaire had to overcome.
By necessity, reversible computing chips operate more slowly than a conventional computer chip, since each logical process needs to be reversed before a component can carry out the next forward computation. Rosini said that Vaire’s design compensates for this slower circuit speed by including many more processing cores than a conventional chip. This kind of parallel processing works particularly well for AI applications which require many similar mathematical operations, such as matrix multiplication, convolutions, and gradient adjustments, to be made across the nodes of a neural network in order to arrive at an output. In fact, the reason GPUs are used for AI is that they contain thousands of parallel processors that can perform these simultaneous calculations (CPUs, by contrast, contain far fewer processing cores and are better suited to computations in which operations need to be performed in a step-by-step sequence).
Many of the earlier attempts at reversible computing were trying to match the performance of CPUs, which meant that the inherently slower clock speed of reversible computing was a significant disadvantage. Vaire’s insight, Rosini said, was to apply a reversible computing architecture to what is essentially a GPU chip design, where there is already a tremendous amount of parallel processing and the slower clock speed matters far less. “I would say the greatest innovation that we brought to the table was not technology. It was simply saying, ‘Hey, this technology can be used to to build GPUs when no one has done this before,’” he said.
Hoping to “fire the starting gun” on a new hardware race
Academic researchers who are aware of Vaire’s work say they have been eagerly awaiting results from tests of its first chip prototypes. Four who spoke to Fortune for this story all said that although it was theoretically possible to conserve almost 100% of the energy in a reversible computing chip, that if Vaire could show even just a large, double-digit percentage power savings for its initial chip, that would be significant. “If they can demonstrate a chip with reversible circuits and reversible clocking [meaning the electricity is recycled] and show an energy efficiency improvement over a conventional CMOS chip performing the same computation — but using a standard circuit style and conventional clocking — then that will be a significant step forward,” the University of Massachusetts’ Anderson said.
What Vaire says it has done so far is show that its resonator component achieves a savings of 50% using computer-based simulations. The entire physical chip, test batches of which are currently in transit from semiconductor fabrication facilities in Asia to engineering teams, has not yet been evaluated. In the past, as with the MIT prototype chips in the 1990s, some of the energy savings of the chips can be eroded by the need to power additional components and the longer circuit pathways that are required for the physical architecture of the reversible circuits.
Rosini said that if Vaire’s chips can show significant energy savings it will “fire the starting gun” on a race among chip companies to build reversible computing chips and optimize the designs to save ever greater amounts of power. He says a company like Vaire just has to demonstrate what’s possible. He compares it the way that few space agencies or companies thought about using reusable rocket components until SpaceX showed it was technically feasible to do so and changed the economics of the whole industry.
He said he is not unaware of the challenges Vaire faces to get companies to accept a new computing paradigm. The entire field of “unconventional computing,” he says, “is a graveyard of companies murdered by Intel over the course of 50 years.” But he said he thinks that the mounting concerns about AI’s energy consumption makes it the perfect moment to try.
