All-to-all reconfigurability with sparse and higher-order Ising machines
Authors:
Srijan Nikhar,
Sidharth Kannan,
Navid Anjum Aadit,
Shuvro Chowdhury,
Kerem Y. Camsari
Abstract:
Domain-specific hardware to solve computationally hard optimization problems has generated tremendous excitement recently. Here, we evaluate probabilistic bit (p-bit) based on Ising Machines (IM) or p-computers with a benchmark combinatorial optimization problem, namely the 3-regular 3-XOR Satisfiability (3R3X). The 3R3X problem has a glassy energy landscape, and it has recently been used to bench…
▽ More
Domain-specific hardware to solve computationally hard optimization problems has generated tremendous excitement recently. Here, we evaluate probabilistic bit (p-bit) based on Ising Machines (IM) or p-computers with a benchmark combinatorial optimization problem, namely the 3-regular 3-XOR Satisfiability (3R3X). The 3R3X problem has a glassy energy landscape, and it has recently been used to benchmark various IMs and other solvers. We introduce a multiplexed architecture where p-computers emulate all-to-all (complete) graph functionality despite being interconnected in sparse networks, enabling a highly parallelized chromatic Gibbs sampling. We implement this architecture in FPGAs and show that p-bit networks running an adaptive version of the powerful parallel tempering algorithm demonstrate competitive algorithmic and prefactor advantages over alternative IMs by D-Wave, Toshiba, and Fujitsu, except a greedy algorithm accelerated on a GPU. We further extend our APT results using higher-order interactions in FPGAs and show that while higher-order interactions lead to prefactor advantages, they do not show any algorithmic scaling advantages for the XORSAT problem, settling an open conjecture. Even though FPGA implementations of p-bits are still not quite as fast as the best possible greedy algorithms implemented in GPUs, scaled magnetic versions of p-computers could lead to orders of magnitude over such algorithms according to experimentally established projections.
△ Less
Submitted 21 May, 2024; v1 submitted 21 November, 2023;
originally announced December 2023.
