Researchers at the National University of Singapore (NUS) have developed spintronic processors integrating 144 tuneable random number generators, offering a potential near-term solution for accelerating complex optimization tasks and reducing energy consumption. The team, led by Professor Yang Hyunsoo from the Department of Electrical and Computer Engineering, reported two advances in Nature Communications detailing probabilistic computing systems built on nanoscale magnetic tunnel junctions. Their parallel magnetic tunnel junction-based processor achieved a 3.2-fold speedup and 58.3 percent energy savings compared to a conventional CPU implementation when solving quadratic assignment problems. “Quantum computing remains an exciting long-term direction, but many optimisation problems need practical solutions today,” said Professor Yang, highlighting the potential for spintronic probabilistic computing to deliver immediate gains in speed and efficiency.
Spintronic Ising Processors Accelerate Quadratic Assignment Problems
This approach bypasses some of the limitations of conventional computing by harnessing the inherent randomness of nanoscale magnetic tunnel junctions to accelerate solutions for complex problems, particularly quadratic assignment problems, a computationally intensive class of optimization challenges. The newly developed processor achieved a 3.2-fold speedup and 58.3 percent energy savings when benchmarked against a standard CPU implementation, demonstrating a substantial improvement in both performance and efficiency. Importantly, the researchers directly compared their spintronic system with D-Wave quantum annealers, revealing a key advantage; while the spintronic processor consistently delivered feasible, high-quality solutions across a full dataset of quadratic assignment problems, the quantum annealers struggled to produce viable results as the problem size increased.
Beyond speed and efficiency, the team also demonstrated scalability, with a second study showcasing a larger probabilistic Ising machine built from 250 spin-transfer-torque magnetic tunnel junctions. This system achieved a 10-fold acceleration for sparsely connected graphs using a cluster parallel update method. Mr Yang Shuhan, a PhD student and first author of both papers, explained their core philosophy: “Instead of treating randomness as a source of error, we use it as a computing resource.” The researchers envision future systems supporting energy-efficient computing platforms for applications including artificial intelligence, logistics, and financial modeling.
The system integrates 144 compact spintronic tuneable random number generators in a massively parallel architecture.
Scalable Magnetic Tunnel Junctions Enable Parallel Probabilistic Computing
These are not theoretical explorations; the work focuses on creating scalable hardware, addressing a critical gap between the promise of emerging technologies and their practical implementation. This processor achieved a 3.2-fold speedup alongside 58.3 percent energy savings. They also showed that employing simulated quantum annealing improved solution quality by 20 times compared to conventional simulated annealing, while simultaneously enhancing robustness against device variability. The team’s future work will focus on scaling up the hardware and exploring chiplet-based architectures, potentially leading to energy-efficient computing platforms applicable to diverse fields including artificial intelligence and logistics.
The processor achieved a 3.2-fold speedup with 58.3 per cent energy savings compared with a central processing unit (CPU) implementation.
