A first-of-its-kind computer can perform calculations using the random “noise” that is inherent in our world. It is built using standard commercial components and could eventually run artificial intelligence programs more efficiently than conventional computers.
In conventional computers, all calculations are reduced to sequences of 1s and 0s, represented as the switching on and off of many tiny switches. However, these computers must contend with random thermodynamic noise, like a piece of a circuit warming up and unexpectedly turning a 0 into a 1. This noise causes errors – but at New York-based start-up and his colleagues have worked out how to build a computer that takes advantage of it instead.
“We are making lemonade out of lemons, using the fact that physical systems are naturally noisy,” says Coles.
Advertisement
He says the team was excited to discover a natural match between the mathematics of a type of artificial intelligence called probabilistic AI and the physics of thermodynamics. This meant that hardware built to recognise and exploit thermodynamic noise could also increase the energy efficiency of probabilistic AI computing.
A thermodynamic computer receives its inputs from the physical environment rather than through a keyboard. For instance, it will detect if one of its components has become warmer. As that warmed component then naturally changes temperature, like cooling down to the more stable ambient temperature, the computer takes advantage of that process to perform a calculation. Because the warm component would change temperature in any case, exploiting this to perform a calculation makes the computer more efficient. The result of the calculation is read out by measuring the properties of the computer’s equilibrium state – the state where it is exceptionally stable and its temperature and voltage no longer change.
The researchers built a prototype thermodynamic computer – which they named the “stochastic processing unit” or SPU – on a circuit board similar to those used in conventional computers. It contains eight interconnected circuits, each of which stores energy in an electric oscillation – a little like an electric version of a swinging pendulum.
To use the SPU, the Normal Computing team exposed it to noisy electrical currents – meaning currents with random, small-amplitude fluctuations rather than a steady flow. The researchers ran different computations by changing these input currents, tuning the circuits so that, for instance, one oscillating “pendulum” has a greater influence over the way the other seven oscillate. They then measured the SPU’s state by measuring a multitude of properties including the currents and voltages of its circuits.
They tested the SPU by successfully running a program that can find the inverse of a so-called “mathematical matrix”, which can be a very challenging calculation. They also ran several programs that are important for building and using generative AI algorithms. The most exciting one is called “uncertainty quantification” and could eventually help “build AI that know how much they don’t know”, says , also at Normal Computing.
“I think it’s not widely appreciated what could be done with thermodynamic computing. And this is a good start,” says at the University of California, San Diego. He says that thermodynamics is at play in conventional computers as well, but there the hardware is built so that all changes in the computer’s state must be directed or forced by the human operator, which can be very energetically costly for uses like AI.
“The fundamental proposition behind thermodynamic computing is that, essentially, if we were less prescriptive in telling the hardware what to do, and let it sort of do these thermodynamics [processes] that are already there, more naturally… then we would get much more capable AI systems and probably ones that are much more energy efficient,” says Hylton.
at Pompeu Fabra University in Spain says that if the SPU were made bigger and could invert matrices very quickly, it would also be useful for computations beyond AI. “Inverting a matrix, that’s important for all numerical computation,” he says. Right now, however, the new computer is a proof-of-principle device. Crucially, the tests involved feeding it electrical currents with in-built fluctuations, rather than having it harness the natural thermodynamic fluctuations in the ambient environment, says Kolchinsky.
But the team at Normal Computing is certain that it can enlarge and improve its current computer to further showcase its potential within five years.
arXiv
