Harnessing Random Noise for Efficient AI Computing

1. New “thermodynamic computer” exploits random physical noise to run calculations instead of fighting against it like conventional computers.
2. Built using commercial hardware, prototype matches probabilistic AI math with thermodynamic physics to increase energy efficiency.
3. Successfully tested on matrix inversion and generative AI programs.
4. Proof-of-concept shows promise but current version still uses controlled noise inputs.
5. Aiming to improve prototype over next 5 years to harness ambient environment noise for efficient AI computing.

A first-of-its-kind computer can perform calculations by harnessing the random “noise” inherent in physical systems. Built using standard commercial components, this prototype “thermodynamic computer” could eventually run artificial intelligence programs more efficiently than conventional computers.

First “Thermodynamic Computer” Exploits Random Noise for Calculations

Rather than representing calculations as sequences of 1s and 0s activated by tiny switches, the thermodynamic computer receives its inputs from, and measures its outputs in, the physical environment. For instance, it may detect a component warming up and then exploit the natural cooling process of that component to perform a calculation. The result is read out by measuring the final stable state after the temperature and voltages settle.

According to Patrick Coles of start-up Normal Computing, there is a natural match between the mathematics of probabilistic AI and the physics of thermodynamics. This means hardware designed to recognize and exploit thermodynamic noise could increase the energy efficiency of probabilistic AI computing.

The researchers built a prototype, dubbed the “stochastic processing unit” (SPU), using circuits that store energy in electric oscillations. To operate the SPU, they exposed it to fluctuating electrical currents and tuned the circuits to influence each other’s oscillations. By measuring currents, voltages and other properties, they successfully ran programs for finding matrix inverses and building generative AI algorithms.

While proof-of-concept, the SPU shows the promise of thermodynamic computing, according to experts. If expanded, it could rapidly invert matrices for broader numerical computations. Currently, the SPU still relies on feeding it fluctuating currents rather than harnessing ambient environment noise. But within five years, Normal Computing aims to further develop thermodynamic computers to showcase their efficiency for AI.