Workshops Archive

Distributed Sparse P-bit Ising Machines for CPU Clusters


Workshop: IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms

Authors: JOSEPH ARNOLD XAVIER (FORSCHUNGSZENTRUM JUELICH), Nan Su (Forschungzentrum Juelich), Morris Riedel (University of Iceland), Gabriele Cavallaro (Forschungszentrum Juelich), David Rabanus (Forschungzentrum Juelich), and EDUARDO SCHATZLE (Forschungszentrum Juelich)

Abstract: Ising computing is a novel paradigm for combi- natorial optimizations. Ising machines built from probabilistic bits (p-bits) update each vertex from a weighted sum over its graph neighbors forming an inherently irregular workload. Distributing the sampler across Message Passing Interface (MPI) processes or ranks lets each rank update its own p-bits while restricting communication to halo p-bits. We characterize this trade-off on 12 MAX-CUT instances using the GPI-2 one- sided communication at 2-8 processes, sweeping synchronization frequency and comparing two primary samplers for the first time: Stochastic Simulated Annealing (SSA) and Probabilistic bit Simulated Annealing (PSA). Poorly partitioned graphs can see a speedup upto 53x from relaxed synchronization. The quality cost, however, is set by density rather than halo/owned ratio. Sparse graphs lose under 1% even at high halo ratios, while dense instances lose 8.6%. PSA yields better solutions in 83% of configurations whereas SSA runs up to 3.4× faster.


Back to IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms Archive Listing

Back to Full Workshop Archive Listing