From Hardware to Algorithms: Probabilistic Computing for Machine Learning, Optimization, and Quantum Simulation

This talk will highlight probabilistic computers as an emerging paradigm for domain-specific computation. Firmly connected to the widely used Markov Chain Monte Carlo algorithms widely used in physics, statistics, and ML, the talk will show how networks of probabilistic bits, or p-bits, in hardware can deliver improvements in time and energy to solutions for ML, optimization, and quantum simulation. Probabilistic computers leverage a physics-inspired architecture with sparse connectivity and asynchronous updates, enabling massive parallelism. Digital implementations in single FPGAs show competitive performance against optimized GPUs/TPUs. Recent efforts with a distributed system of multiple FPGAs creates the “illusion” of a single, more powerful system, achieving near-linear speedup with minimal communication overhead. Beyond digital CMOS, magnetic nanodevices offer intrinsic randomness, replacing thousands of transistors per p-bit and reducing energy per operation. Our ongoing efforts aim to integrate these devices into energy-efficient CMOS+X systems. Comparisons with quantum computers, GPUs/TPUs, and coupled oscillators will illustrate how probabilistic computers combined with tailored algorithms could achieve GPU-like impact and enable new applications. Speaker(s): , Kerem 5000 Forbes Ave, Bosch Spark Conference Room, Scott Hall 5201, Pittsburgh, Pennsylvania, United States, 15213