Probabilistic computing treats randomness as a resource rather than a nuisance, using controlled noise to explore the vast solution spaces of combinatorial optimisation. A team in the NUS Department of Electrical and Computer Engineering has built the hardware to do this in silicon-compatible form, using magnetic tunnel junctions: nanoscale devices whose switching behaviour generates randomness that can be tuned electrically. Two studies in Nature Communications report the resulting machines.
The first describes a parallel processor incorporating 144 spintronic random number generators, aimed at quadratic assignment problems. It achieved a 3.2-fold speedup and 58.3 per cent energy savings against a processor-based implementation, and returned feasible, high-quality solutions consistently in regimes where quantum annealers struggled as problem size grew. That last comparison matters, because quadratic assignment is precisely the class of problem often cited as a near-term quantum application.
The second study presents a probabilistic Ising machine built from 250 spin-transfer-torque magnetic tunnel junctions. A cluster parallel update method delivered a tenfold speedup on sparsely connected graphs, and simulated quantum annealing improved solution quality twentyfold over conventional simulated annealing. Taken together the two systems demonstrate that the approach scales with device count rather than depending on a single exotic component.
The strategic argument is one of timing. Practical quantum advantage for optimisation remains distant, while spintronic hardware builds on magnetic memory technology that foundries already understand. If probabilistic processors can be fabricated alongside conventional logic, then logistics scheduling, financial modelling and electronic design automation acquire a near-term accelerator without waiting for cryogenic quantum systems to mature.