CIMs employ an optical storage-ring memory format to enable massive connectivity.
CIMs are hybrid computers that utilize both electronic and optical logic. The primary memory register of the CIM architecture comprises a set of optical pulses (whose coherent amplitudes represent CIM data) stored via ballistic propagation around a free-space resonator or optical fiber ring. The amplitude of each data pulse is measured once per round trip of the storage ring, and the full vector of amplitudes is stored and continually updated in an FPGA data processing engine. The FPGA engine drives a laser-modulator subsystem that applies incremental perturbations to the optical data amplitudes once per round trip, which effectively implements a desired set of Ising couplings among the dynamic variables. Signal restoration to compensate for propagation losses around the ring is implemented via coherent parametric amplification in a nonlinear crystal; the pump strength applied to this crystal (and therefore the round-trip “gain” of the optical storage ring) is increased during the computation, resulting eventually in a discontinuous threshold phenomenon that realizes a final optical logic operation. The current hardware layout based on an optical storage ring and stationary FPGA-based data processing engine has supported rapid scaling of prototype CIMs with all-to-all connectivity of the optically represented Ising spins. Can we preserve the advantages of this memory scheme in next-generation nanophotonic hardware platforms? Can it be improved upon? Could versions of it provide advantages in other unconventional computational architectures?
P. L. McMahon et al., “A fully programmable 100-spin coherent Ising machine with all-to-all connections,” Science 354, 614 (2016).
A. Marandi et al., “Network of time-multiplexed optical parametric oscillators as a coherent Ising machine,” Nature Photonics 8, 937 (2014).