Hybrid Photonic Iterative Solver for Large-Scale Systems of Equations

Abstract

The University of Central Florida invention enables a photonic iterative solver (PIS) for many large-scale linear inverse and optimization science and engineering problems. Example applications include matrix inversion, image reconstruction, high-dimensional inverse problems, numerical eigen problems, and ab initio ray-tracing engines.

Conventional iterative solvers with floating-point computing units are slow and inefficient at handling large-scale or high-dimensional data. Though systems using a fixed-point data format have the potential to train large-scale neural networks, fixed-point is not suitable for iterative solvers that require precision on the output. The limited precision can stagnate the iterative solver before reaching the optimal solution. Also, existing fixed-point iterative solvers in analog computing hardware have no error management mechanisms to resolve the error stagnation problem.

The UCF solution offers a hybrid analog and digital computational system that uses a dynamic fixed-point format to reach the same precision level as a conventional floating-point iterative solver. Compatible with the latest hardware advancements in high-performance computing, the invention achieves high speed and energy efficiency with fixed-point, highly parallelized architecture.

Technical Details: The UCF invention comprises a system, computer program product, and method of creating a hybrid analog and digital computational system. It incorporates a residual iterative algorithm to solve the set of solution values for equations. The residual iterative algorithm includes an outer update loop computed using a digital computing device with residue values initially set to a first initial value and a set of solution update values set to a second initial value.

The residual iterative algorithm also includes an inner residual loop that is iteratively computed using an analog accelerator until one or more inner residual loop stopping criteria are met and return a set of solution update values. Next, the system uses the new values to update the set of residue values and a range of the next set of solution update values, thereby adjusting the computational precision of the inner residual loop.

Partnering Opportunity: The research team is looking for partners to develop the technology further for commercialization.

Stage of Development: Prototype available.

Benefit

Market Application