Probabilistic Computing with P-Bits (arxiv.org)

The paper proposes a new class of computing — “p-computers” built from p-bits, classical stochastic bits that flip between 0 and 1 with controllable probabilities. Unlike deterministic bits or quantum qubits (whose amplitudes require coherence), p-bits intentionally encode and manipulate probability distributions in hardware. The authors present a generic architecture for networks of p-bits, argue that such devices can be compact and energy-efficient, and emulate systems with thousands of p-bits to demonstrate substantial speedups on randomized workloads. Significance for AI/ML lies in mapping core probabilistic tasks directly onto hardware: sampling, Bayesian inference, optimization, Ising-model solvers, and quantum Monte Carlo all benefit because p-bits natively generate correlated stochastic samples used by Monte Carlo and Boltzmann-like algorithms. By sidestepping quantum coherence requirements while still operating on probability distributions, p-computers offer a practical, scalable accelerator for stochastic algorithms central to machine learning and combinatorial optimization. The work suggests a promising, hardware-oriented pathway to faster, lower-power probabilistic computation that could complement existing accelerators for training and inference.