An AI energy breakthrough is claimed by Extropic's chip

Thermodynamic sampling units (TSU), a revolutionary chip architecture that does probability calculations rather than conventional processing, were recently presented by Extropic. The company claims the technology can run AI models with 10,000 times less energy than existing GPUs. The specifics: Extropic's chips trade accuracy for reduced energy consumption by directly generating likely solutions rather than working step-by-step like conventional GPUs. In addition to providing open-source tools for researchers to test the methodology, the business sent its initial development kit to weather companies and AI laboratories. Next year, Extropic will release its Z-1 chip, which is intended to run a novel diffusion model that generates images and videos by gradually eliminating noise. Extropic, which was founded by former Google quantum researchers, thinks that unless the industry gives up on its existing chip designs, energy constraints would hinder the advancement of AI. Extropic has been hinting at a paradigm shift in AI hardware for years under the direction of Guillaume Verdon, who was identified as the well-known @BasedBeffJezos on X. Today appears to be the first significant (public) step towards that vision. The trillion-dollar question will be whether the technology can be applied to real-world, production-ready systems at scale.