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.
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.
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.
