Extropic AI Targets 10,000x Energy Efficiency with New Thermodynamic AI Chips

Extropic AI, a startup pioneering thermodynamic computing, has recently unveiled a novel chip architecture aiming to deliver significantly greater energy efficiency for artificial intelligence workloads. The company claims its new Thermodynamic Sampling Units (TSUs) could perform certain generative AI tasks with up to 10,000 times less energy than conventional GPUs. This development has sparked discussion within the tech community, with prominent industry figure Naveen Rao publicly defending the innovative approach.

Unlike traditional GPUs that are highly optimized for deterministic operations, Extropic's TSUs are designed to harness thermodynamic electron fluctuations to model probabilities directly. This approach moves away from fighting inherent electronic noise in silicon, instead leveraging it for probabilistic AI workloads. The company describes its hardware as inherently probabilistic, a "perfect fit" for modern AI.

Naveen Rao, a respected voice in the AI and semiconductor industry, expressed his support for Extropic AI's direction amidst what he perceived as undue criticism. > "I’m not getting the reason for all the hate. I’m obviously missing context on these conversations, and maybe it’s just personal. But it looks like @Extropic_AI did build something interesting," Rao stated in a recent social media post. He further suggested that "new hardware that is efficient in different operations might be used to go in new directions" for AI, questioning the "backlash to that notion."

Founded in 2022 by Guillaume Verdon, a former Google quantum computing researcher, Extropic recently secured $14.1 million in seed funding. This round was led by Kindred Ventures, with participation in notable investors including Garry Tan and Balaji Srinivasan. The company's mission is to build radically energy-efficient computing hardware, directly addressing the vast energy demands of the current AI industry.

Verdon, also known by his controversial online persona "Based Beff Jezos" and associated with "effective accelerationism" (e/acc), has been a vocal critic of traditional AI safety movements. This background, advocating for unrestrained technological progress, may contribute to the "backlash" Rao alluded to. Extropic's computing paradigm fundamentally contrasts with quantum computing by embracing noise as an asset rather than a liability, seeking to build a physics-based system.

Extropic has already shared its first working chip, XTR-0, with a select group of partners, including frontier AI labs and government representatives. The company is also developing an open-source Python library, THRML, to enable developers to experiment with thermodynamic algorithms. This novel paradigm seeks to offer a less costly and more sustainable alternative to the increasingly power-hungry data centers driving current AI development.