Programmable Probabilistic Computer with 1M p-bits
4 days ago
- #probabilistic computing
- #distributed sampling
- #Ising machine
- Researchers developed a probabilistic computer with 1 million p-bits by networking FPGAs, overcoming single-chip limitations.
- The machine performs Gibbs sampling at over a trillion flips per second, with all coupling weights stored in local on-chip memory.
- Boundary information exchange between devices is minimal, using only 1-bit states.
- A key finding is the timing ratio η = f_comm / f_p-bit, which determines boundary refresh frequency for distributed sampling to mimic monolithic systems.
- Experiments on 3D Edwards-Anderson spin glasses show a threshold for η above which performance matches a GPU reference; below it, a tradeoff between throughput and accuracy emerges.
- A theoretical cluster mean-field model confirms the universality of this tradeoff in partitioned stochastic dynamics.
- The platform is programmable and demonstrated on problems like spin glasses, Max-Cut, and Boolean satisfiability.