Probabilistic Processing-in-Memory Computation Enabled by Heterogeneous Magnetic Tunnel Junctions

Current processing-in-memory (PIM) technologies face challenges from large peripheral overhead, limited precision, and low memory density, hindering their adoption for general-purpose computing. In this work, we suggest a probabilistic PIM (p-PIM) computer that harnesses intrinsic stochasticity of magnetic tunnel junctions (MTJs) for direct probability sampling without explicit value computation. By stacking stochastic MTJs (sMTJs) on conventional MRAM cells and incorporating simple logic gates, the design not only supports efficient matrix vector multiplication operations but also demonstrates fundamental advantages in probabilistic algorithms. Using a heterogeneous MTJ-FPGA prototype, we validate the p-PIM computer's capabilities and advantages through machine learning inference, Ising machine implementation, and sequential Bayesian inference where traditional PIM architectures optimized for matrix operations struggle with the inherently sequential updates. The combination of probabilistic computing capability, high memory density, and CMOS compatibility makes our p-PIM architecture a promising platform for both conventional and probabilistic computing applications.