exotic, a startup developing an exotic up-to-date type of computer chip that supports probabilistic bits has produced its first working hardware, along with proof that more advanced systems can perform useful tasks in artificial intelligence and scientific research.
The startup’s chips work in a fundamentally different way than those from Nvidia, AMD and others, and when scaled up, they are expected to be thousands of times more energy competent. With the influx of AI companies billions of dollars to build data centersan entirely up-to-date approach could provide a much cheaper alternative to a wide range of conventional chips.
Extropic calls its processors thermodynamic sampling units (TSUs), as opposed to central processing units (CPUs) or graphics processing units (GPUs). TSUs operate silicon components to harness and shape the thermodynamic fluctuations of electrons to model the probabilities of various sophisticated systems, such as the weather, or artificial intelligence models capable of generating images, text or videos.
Extropic’s first working chip has now been made available to a handful of partners, including pioneering AI labs, weather modeling startups, and representatives from several governments. (Extropic declined to name names).
“This allows all kinds of developers to kick the tires,” says Extropic CEO Guillaume Verdon, who rose to fame in the tech world as a colorful and sometimes controversial internet figure called Based on Beff Jezos and a up-to-date techno philosophy known as effective accelerationism or e/acc before founding a startup. Verdon and his co-founder, Trevor McCourt, who is Extropic’s chief technology officer, previously worked on quantum computing at Google before taking the novel approach to computing.
One of those now testing the up-to-date equipment is Johan Mathe, CEO of Atmo, a startup that uses artificial intelligence models that can make predictions at a higher resolution than is otherwise possible. Its clients include the Department of Defense. Mathe says Extropic tokens should be able to calculate the odds of different weather conditions much more effectively.
Extropic also releases software called TRHML that allows you to simulate the behavior of the Extropic chip on the GPU. Mathe used this software as well as a real chip. “I was able to run a few p-bits and see if they behave as they should,” says Mathe.
The company’s hardware, called the XTR-0, consists of a field-programmable gate array (FPGA) chip that can be reconfigured for a variety of tasks, combined with its first two probabilistic chips, the X-0, each of which contains a handful of p-bits.
Instead of conventional bits corresponding to 1 or 0, the up-to-date chip contains probabilistic bits, or p-bits, that determine the uncertainty of the model. Despite its circumscribed scale, the up-to-date chip demonstrates the potential of the company’s up-to-date approach.
“We have primitive machine learning that is much more efficient than matrix multiplication,” McCourt says. “The question is how to build something at the scale of ChatGPT or Midjourney.”