Save StorySave this storySave StorySave this story
Extropic, a startup developing an exotic new kind of computer chip that handles probabilistic bits, has produced its first working hardware along with proof that more advanced systems will tackle useful tasks in artificial intelligence and scientific research.
The startup’s chips work in a fundamentally different way to chips from Nvidia, AMD, and others, and promise to be thousands of times more energy efficient when scaled up. With AI companies pouring billions of dollars into building datacenters, a completely new approach could offer a far less costly alternative to vast arrays of conventional chips.
Extropic calls its processors thermodynamic sampling units, or TSUs, as opposed to central processing units (CPUs) or graphics processing units (GPUs). TSUs use silicon components to harness thermodynamic electron fluctuations, shaping them to model probabilities of various complex systems, such as the weather, or AI models capable of generating images, text, or videos.
The first working Extropic chip has now been shared with a handful of partners including frontier AI labs, startups working on weather modelling, and representatives from several governments. (Extropic has declined to provide names.)
“This allows all sorts of developers to kick the tires,” says Extropic CEO Guillaume Verdon, who gained notoriety within the tech world as a colorful and sometimes controversial online persona called Based Beff Jezos and a new techno philosophy known as effective accelerationism or e/acc before founding the startup. Verdon and his cofounder Trever McCourt, who is Extropic’s CTO, previously worked on quantum computing at Google before pursuing their novel computing approach.
One of those who is now testing the new hardware is Johan Mathe, CEO of Atmo, a startup that uses AI models forecast with higher resolution than is normally possible for customers including the Department of Defense. Mathe says that Extropic’s chips should make it possible to calculate the odds of different weather conditions far more efficiently than is normally possible.
Extropic is also releasing software called TRHML that makes it possible to simulate the behavior of an Extropic chip on a GPU. Mathe has used this software as well as the real chip. “I was able to run a few p-bits and see that they behave the way they are supposed to,” Mathe says.
The company’s hardware, called XTR-0, consists of a field programmable gate array (FPGA) chip, which can be reconfigured for different tasks, combined with two of its first probabilist chip, X-0, each of which contains a handful of qubits.
The XTR-0.
Single daughterboard.
Instead of conventional bits corresponding to either 1 or a 0, the new chip features probabilistic bits or p-bits that model uncertainty. Although limited in scale, the new chip demonstrates the potential of the company’s new approach.
“We have a machine learning primitive that is far more efficient than matrix multiplication,” McCourt says. “The question is how do you build something on the scale of ChatGPT or Midjourney.”
In a paper posted to arXiv, the company lays out how a larger chip with thousands of p-bits, which it claims it can to deliver next year, could be used to create a new kind of diffusion model—an important type of model that is used to generate images and videos and to guide robot’s actions.
“It could be a huge win,” Mathe says of the forthcoming chip, dubbed Z-1, which Extropic says will have 250,000 p-bits.
“Their approach to the physics of information processing could prove transformative over the next decade, particularly as conventional transistor scaling hits fundamental limits,” adds Vincent Weisser, CEO of Prime Intellect, a startup working on distributed AI approaches. “If scaled practically, it could deliver orders-of-magnitude improvements in energy efficiency and density, critical for workloads where energy per operation is a bottleneck.”
Verdon and McCourt argue that the incredible amount of money being poured into building AI datacenters ignores the incredible energy requirements that such a boom would entail. “Even if we have a one percent chance of success—and we think it’s much higher than that—it’s worth trying,” McCourt says.
