Researchers have introduced the concept of a thermodynamic computer—a new type of computing architecture that could significantly reduce the energy consumption of artificial intelligence systems. This was reported by Quantum Insider.
A team from Extropic and the Massachusetts Institute of Technology believes their approach could make certain AI tasks up to 10,000 times more energy-efficient compared to traditional computing.
Leveraging Noise Instead of Fighting It
Modern processors, including GPUs used to train and run large language models, expend significant resources on precise deterministic calculations. Physical noise and thermal fluctuations are typically seen as disturbances that need to be minimized.
The authors propose a contrary approach. Instead of suppressing random thermal processes, they suggest utilizing them as part of the calculations. This principle is termed thermodynamic computing.
According to the researchers, many AI tasks—such as finding the most probable answer or optimal solution—are inherently probabilistic. Therefore, a computational system that harnesses random physical processes could potentially perform these tasks much more efficiently than classical processors.
Tackling One of AI's Major Challenges
The interest in such architectures stems from the growing energy demands of the AI industry. Major tech companies are investing billions in building data centers, and the demand for electricity to train and operate modern models continues to rise rapidly.
If the proposed architecture proves viable, it could not only reduce energy costs but also lower the operational expenses of AI infrastructure, decreasing the need for expensive computing clusters.
Still a Long Way to Practical Application
Currently, this is fundamental research rather than a ready-to-use processor. The authors have presented the architecture and simulation results demonstrating the advantages of the new approach for specific classes of tasks. It may take years before commercial chips based on thermodynamic computing principles are available.
Nonetheless, this work reflects the growing interest in alternative computing architectures within the industry. As the scale of AI models continues to expand, there is increasing focus not only on their capabilities but also on the cost of computations. In this context, finding ways to drastically reduce energy consumption is becoming a key area of research, alongside the development of quantum and neuromorphic computers.
In May, Amazon implemented a new architecture for data center networks that accelerates data transfer while also reducing energy consumption.