Reversible computing, a concept first introduced by Rolf Landauer in the 1960s, has gained significant attention in recent years due to its potential to revolutionize the field of artificial intelligence (AI). The idea is simple yet profound: instead of deleting data, reversible computers would store all information and use it to run calculations forward, then backward.
The concept of irreversible computing was born out of a fundamental principle of thermodynamics. When a computer deletes data, electrons in its chips go from known paths to unknown ones, effectively losing energy as heat. This loss is inevitable and makes traditional computing less efficient.
However, Landauer’s discovery also opened up an alternative approach: reversible computing. Charles Bennett proposed the idea of “uncomputation” – running each calculation forward, storing the result, then running it backward without deleting any data. While this method takes twice as long, researchers have since made improvements by using slightly more memory.
The breakthrough came in 2022 when Hannah Earley demonstrated a rigorous account of reversible computing’s efficiency. She found that reversible computers emit significantly less heat than traditional ones. This is crucial for AI applications, where computations are often run in parallel, creating an opportunity for reversible computing to shine.
By running reversible chips more slowly but using more of them to compensate, the energy-saving advantage wins out over the slower processing time. Moreover, this approach can reduce cooling requirements and enable the stacking of chips closer together, saving on space, materials, and time spent on data transfer.
Researchers, including Torben Ægidius Mogensen, are optimistic about the potential for reversible computing in practice. The development of commercial versions of reversible chips is underway, with investors taking notice. As we move forward, it’s exciting to think that reversible computing may finally be put into action and revolutionize the way AI systems process information.
Source: https://www.quantamagazine.org/how-can-ai-researchers-save-energy-by-going-backward-20250530