Research: Unconventional Computing
From reversible to thermodynamic computing — a pursuit for energy efficiency
Date: 2026-09-22
In the 1960s~1980s, physicists and computer scientists explored the fundamental physical limits of computation through the lens of thermodynamics using, concepts such as heat and entropy. One result was reversible computing (RC), which seeks to avoid unnecessary energy loss due to information erasure. Proposed implementation included reversible gates such as the Toffoli and Fredkin gates which later became important primitives in quantum circuit designs. For several decades, however, RC remained peripheral as Moore’s law and Dennard’s scaling continued driving improvement in performance and energy efficiency of conventional silicon computing.
More recently, massive energy and artificial intelligence build-outs have raised concerns over energy use and sustainability, renewing interest in alternative computation paradigm. Building on the concepts of thermodynamics of computation and advances in non-equilibrium information thermodynamics, thermodynamics computing (TC) has emerged as one such approach. Unlike early RC, TC embraces thermal noise, fluctuations and stochasticity rather than suppressing them. This offers several potential advantages. First, TC can reduce the energy spent maintaining highly stable, deterministic states. Second, these fluctuations perturb the system and help it explore different state. The resulting non-equilibrium dynamics can then perform useful computation through processes such as relaxation (annealing), stabilization (sampling) and adaptation (learning).
More broadly, the pursuit of a greater computing efficiency extends beyond TC and RC to several overlapping unconventional computing paradigms. Near-zero energy computation builds on the principles of early reversible computing while acknowledging that perfect reversibility is practically impossible. Instead, these systems recover and reuse part of the energy that would otherwise be dissipated by driving electronic components more adiabatically and designing computing elements with some degree of reversibility, reducing the net energy required for computation. Probabilistic computing (PC) is perhaps the closest neighbor to TC, both exploiting the dynamics of networks of interacting stochastic elements. PC, however, often abstract away much of the microscopic dynamics into standardized probabilistic elements such as probabilistic bits, and utilize their collective statistics for computation.
We investigate the strengths and weaknesses of these technologies, elucidate the complex overlapping relationships between these paradigms and explore how they may complement one another to create a solution that is greater than the sum of the individual approaches.
Probabilistic computing — a new energy efficient solution
Date: 2026-09-17
Modern artificial intelligence (AI) is advancing at a breathtaking pace, arguably unmatched since the industrial revolution. Building on the transformer architecture, larger models and training datasets, and increasingly sophisticated post-training and inference techniques such as reinforcement learning, inference-time reasoning and tool use, AI systems are now capable of tasks that until recently would have been considered science fiction. These gains in capability drive wider adoption, which in turn incentivizes further scaling and improvement in a self-reinforcing feedback loop.
To support this growth, graphics processing units have so far scaled remarkably well with the demands of AI software. Increasingly, however, the supporting infrastructure is struggling to keep pace. Memory shortages, pressure on power grids, water consumption and noise pollution associated with large data centres are now recurring concerns. Part of the problem lies in a mismatch between conventional computing architectures and the structure of certain computational problems, resulting in substantial computational overhead.
Bayesian inference and combinatorial optimization are some examples of problem families that can be poorly matched to conventional deterministic computing systems, often requiring extensive sampling or brute-force searches over enormous solution spaces. Researchers have therefore explored more physics-aware computing approaches that exploit the inherent structure of these problems more efficiently. One such approach is probabilistic computing, which does not merely tolerate stochasticity but incorporates probability directly into the computational model. This can enable algorithms such as Boltzmann-machine sampling and parallel tempering to explore complex probability distributions and optimization landscapes more efficiently, potentially reducing both computational and energy costs.
Interest in these alternative computing architectures has grown rapidly. We explore the steps needed to move probabilistic computing from an academic novelty toward an impactful technology, with emphasis on the role of academic–industry partnerships, which we believe is key to the scaling of this technology.