Programmable Probabilistic Computer with 1M p-bits
5 hours ago
- #Distributed systems
- #Probabilistic computing
- #Ising machines
- A programmable probabilistic computer with one million p-bits is realized by networking FPGAs, breaking the single-chip limit for Ising model acceleration.
- The machine achieves Gibbs sampling at over a trillion flips per second while storing all coupling weights locally, exchanging only 1-bit boundary states between devices.
- A key finding is that boundary information refresh frequency, characterized by the ratio eta = f_comm/f_p-bit, determines if the distributed system matches a monolithic reference.
- For 3D Edwards-Anderson spin glasses, exceeding a topology-dependent eta threshold ensures accurate sampling; below it, a throughput-accuracy tradeoff occurs with reduced decay exponent in residual energy.
- A cluster mean-field model theoretically validates this tradeoff as universal for partitioned stochastic dynamics, providing a design rule for scaling probabilistic computers.
- The platform is demonstrated across spin glasses, Max-Cut, and Boolean satisfiability problems.