A New Concept Computer Specialized in Probabilistic Computation... Will It Usher in a New Era of AI? [Tech Talk]

A New Probabilistic Computer
Closely Mirrors AI Training and Inference
"10,000 Times More Efficient for AI Tasks Than Nvidia GPUs"

Computers are machines that operate using just two numbers: 0 and 1. From the earliest computers to today’s artificial intelligence (AI) semiconductors, this rule has never been broken. However, a recent American startup claims to have invented a new kind of computer that probabilistically generates 0s and 1s to directly perform calculations for AI models. The name of this computer is the Thermodynamic Sampling Unit (TSU). It is a completely new concept of computer developed with the sole purpose of implementing AI.

Physicist Turned AI Researcher... Introduces a Paradigm-Shifting Chip

The TSU was developed by Extropic, a startup founded by Canadian mathematician and physicist Guillaume Verdon. According to the official website, Verdon is an author of research papers on Google’s quantum AI technology "TensorFlow Quantum" and has devoted himself to AI research, including founding a quantum computing startup. However, he encountered the limitation that existing computer semiconductors could not properly implement AI. As a result, he canceled all his projects, founded Extropic in 2022, and focused on developing a new concept of computer.

XTR-0, the world's first probabilistic computer semiconductor introduced by Extropic. Extropic

After three years of silence, Extropic unveiled the TSU earlier this month. On the surface, it looks similar to conventional computer chips such as central processing units (CPUs) and graphics processing units (GPUs). Like other semiconductors, it is made of silicon and assembled with other components on an electronic board. However, Extropic claims in its technical documentation that "the TSU performs AI tasks 10,000 times more efficiently than Nvidia’s GPU." How is the TSU able to achieve such performance?

Deterministic Computer vs. Probabilistic Computer

It is often said that "computers think in 0s and 1s." This is related to the computer’s core component, the semiconductor transistor. The state in which current flows through a transistor is treated as on (1), and the state in which it does not flow is treated as off (0). Logic is implemented based on the state of the transistors. For example, when we move a mouse or press a keyboard, those commands are converted into countless 0s and 1s so the computer can recognize them.

The reason only 0s and 1s are used is that it allows the computer to receive the minute signals of electric current without error. Regardless of the strength of the current, if it flows, it is judged as 1; if it does not, it is judged as 0. This minimizes mechanical errors and produces results exactly as predicted.

An analogy of the operation methods of a conventional computer semiconductor (above) and a thermodynamic sampling unit (TSU). While the semiconductor outputs fixed 0s and 1s depending on whether current flows through the transistor, the TSU probabilistically outputs 0s and 1s based on the wave nature of electrons. Online community, Extropic Capture

In contrast, the TSU produces values not based on current signals but on the vibrations, or waves, of electrons. While it still outputs only 0s and 1s like traditional semiconductors, which value it outputs is purely random. Extropic defines the TSU’s operating method as a "probabilistic algorithm."

TSU Closely Resembles AI Operation

Computers are designed to perform calculations based on 0s and 1s, so if these values were generated randomly, the computer would not function. In fact, the TSU cannot run ordinary software. Instead, it treats the randomly generated calculation results as "samples." By collecting as many samples as possible, it recognizes large-scale patterns.

This feature is very similar to neural network AI. AI trains on vast amounts of data and extracts appropriate patterns from that data. When we ask an AI chatbot a specific question, the chatbot selects and combines the words or images that are most likely to match the patterns it has already learned, and presents them to us.

Whereas conventional computers transform data into astronomical matrix multiplications and calculate each one when training AI models, a computer that implements probabilistic patterns can skip the complex calculation process and directly output results. Verdon explained, "AI algorithms are ultimately nothing more than probabilistically distributed matrix multiplications," and "while conventional computers performed countless matrix multiplications directly, the TSU skips the matrix multiplication and samples the probability distribution itself." He claims that "it is possible to achieve the best AI performance per watt (W) of power consumption."

To Become Mainstream, Execution Environment Must Be Established

Nvidia is also a company that established the concept of the graphics processing unit (GPU), but it faced numerous hardships before the GPU boom. Photo by Reuters Yonhap News

So, will the TSU surpass Nvidia’s GPU and become the dominant AI semiconductor in the future? For now, it remains only a possibility. Even if the TSU shows superiority in tasks requiring probabilistic algorithms, such as AI, nuclear simulations, and energy management, its applications are limited because a proper execution environment has not yet been established. It cannot currently surpass the ecosystem and versatility that GPUs have built in the AI market.

Verdon also admitted, "We have succeeded in developing the TSU, but we need more help to reach a practical stage." Extropic is currently seeking collaboration with researchers in physics, biology, AI, and energy who perform probabilistic computations to build the necessary execution environment. Recently, a proposal was posted on the official website stating, "We want to collaborate with companies that wish to perform large-scale probabilistic tasks to develop system applications."

When Nvidia launched the world’s first GPU, the GeForce 256, it did not succeed from the start. CEO Jensen Huang personally led marketing and sales efforts, and after decades of developing various programs that could run on GPUs, Nvidia was able to claim the title of winner in the AI era. How probabilistic computers will change the future now depends on Extropic’s next moves.