New computer architecture aims to make AI more energy-efficient

Artificial intelligence is consuming increasing amounts of energy. Major technology companies are investing billions in new data centers to train and run AI models. Researchers now believe they have found a way to drastically reduce that energy consumption.

A team of researchers from Extropic and the Massachusetts Institute of Technology presented a new concept for a thermodynamic computer. According to the researchers, this could make certain AI computations up to 10,000 times more energy-efficient than with current computers.

The opposite approach

Most modern chips, such as the graphics processors used for large language models, are designed to perform extremely precise calculations. In doing so, heat and random disturbances are suppressed as much as possible.

The researchers have opted for the opposite approach. Instead of combating noise and thermal fluctuations, they aim to use these natural processes as part of the computation itself. This method is known as thermodynamic computing.

AI applications

According to the scientists, this approach is well-suited to many AI applications. After all, artificial intelligence often does not seek a single exact answer, but rather the most likely solution. Random processes in nature could perform such calculations more efficiently than traditional chips.

Interest in alternative computer architectures is growing

Interest in alternative computer architectures is growing rapidly as AI’s energy consumption continues to rise. A more energy-efficient form of computing would not only reduce data centers’ power consumption but also lower the costs of the necessary infrastructure.

Basic research

For now, this is still basic research. The researchers have described the architecture and run simulations, but a commercial chip based on this technology seems years away.

Nevertheless, the research shows that the search for new forms of computer technology is in full swing. In addition to quantum computers and neuromorphic chips, thermodynamic computing is also receiving increasing attention as a potential solution to AI’s growing energy problem.