250 Magnetic Tunnel Junctions-Based Probabilistic Ising Machine

Abstract

In combinatorial optimization, probabilistic Ising machines (PIMs) have gained significant attention for their acceleration of Monte Carlo sampling with the potential to reduce time-to-solution in finding approximate ground states. However, to be viable in real applications, further improvements in scalability and energy efficiency are necessary. One of the promising paths toward achieving this objective is the development of a co-design approach combining different technology layers including device, circuits and algorithms. Here, we experimentally demonstrate a fully connected PIM architecture based on 250 spin-transfer torque magnetic tunnel junctions (STT-MTJs), interfaced with an FPGA. Our computing approach integrates STT-MTJ-based tunable true random number generators with advanced annealing techniques, enabling the solution of problems with any topology and size. For sparsely connected graphs, the massive parallel architecture of our PIM enables a cluster parallel update method that overcomes the serial limitations of Gibbs sampling, leading to a 10 times acceleration without hardware changes. Furthermore, we prove experimentally that the simulated quantum annealing boosts solution quality 20 times over conventional simulated annealing while also increasing robustness to MTJ variability. Short pulse switching measurements indicate that STT-MTJ-based PIMs can potentially be 10 times faster and 10 times more energy-efficient than graphic processing units, which paves the way for future large-scale, high-performance, and energy-efficient unconventional computing hardware implementations.