Thermodynamic computing may revolutionize the way images are generated, offering a much lower energy cost compared to current digital hardware. This discovery is the fruit of research from Stephen Whitelam—a staff scientist at the Lawrence Berkeley National Laboratory. He and his colleague’s research was published Jan. 10 in the journal Nature Communications.
Normal Computing, a New York City-based startup, has developed a remarkable prototype chip. This chip contains eight resonators, or waves of sound, that are physically linked with purpose-built couplers. This chip is a platform for the new thermodynamic computing approach. The researchers simulated this chip on traditional computers. On January 20, they published their results in Physical Review Letters.
Whitelam goes on to explain that the training process for this kind of thermodynamic computer would look immensely different than that of the original one. As he put it, “This study demonstrates that we can design specialized hardware to perform complex types of machine learnings (e.g., image generation) and so it can do so at much lower energy expenses than what we do today.
Even with these exciting advances, Whitelam concedes that thermodynamic computers are still quite basic compared to today’s digital neural networks. Yet he knew the monumental fight that lay before him. We don’t know how to design a thermodynamic computer that would be as proficient at image generation as, for example, DALL-E,” he added, highlighting the distance between today’s technology and what’s feasible down the road.
Whitelam knows there’s a long way to go. We still have a long way to go to build the hardware required for a complete rollout of this technology. He added that researchers and engineers in the field need to work out how to construct the hardware. We hope this comment lays out a pretty simple and obvious road forward for their work.