IN Brief:
- LightSolver and HLRS evaluated the Laser Processing Unit architecture against GPU algorithms for sparse linear systems.
- The study used an LPU emulator and projected acceleration ranging from around 40 times to more than 80,000 times across selected benchmarks.
- The architecture uses a programmable optical resonator to perform iterative computation while retaining state within the optical system.
LightSolver and the High-Performance Computing Center Stuttgart have published research examining the use of optical processing to accelerate large sparse systems of linear equations, a workload that underpins many scientific and engineering simulations.
The study compares LightSolver’s Laser Processing Unit architecture with established iterative algorithms running on a GPU. The researchers used an emulator of the LPU rather than completed production hardware, allowing the architecture to be evaluated across a series of mathematical problems before the same performance can be demonstrated on a physical system.
Projected time-to-solution improvements range from about 40 times to more than 80,000 times across the selected benchmarks and reference algorithms. The range reflects the characteristics of individual problems and does not represent a single performance multiplier that can be applied across high-performance computing workloads.
Sparse linear systems occur in computational fluid dynamics, structural mechanics, electromagnetics, molecular modelling, and other applications where very large matrices contain relatively few non-zero values. Repeatedly solving those systems can consume a substantial proportion of overall simulation time, particularly as model size increases.
LightSolver’s architecture uses a programmable optical resonator in which light circulates around a closed path. A spatial light modulator applies programmed operations during successive passes, allowing the optical state to evolve through repeated iterations while information remains within the resonator.
The approach differs from conventional digital processing, where numerical values are repeatedly moved between compute units, registers, caches, and memory. Keeping computation and state within the optical system is intended to reduce the memory movement associated with highly iterative calculations.
The LPU is not designed as a general replacement for CPUs or GPUs. LightSolver is targeting mathematical workloads that map onto the resonator architecture, while conventional processors continue to handle control, scheduling, data preparation, storage, communications, and software that does not suit the optical engine.
That points towards a heterogeneous computing model similar in principle to current HPC systems, where CPUs, GPUs, and specialist accelerators each handle different classes of work. An optical processor would become another resource in that system rather than a standalone substitute for the existing compute stack.
LightSolver states that its LPU architecture can provide the equivalent of around 200TB/s of internal optical memory bandwidth and about 1,000 TOPS while consuming approximately 100W. Those figures describe the company’s architecture and are separate from the new emulator-based GPU comparison.
Physical implementation introduces additional variables that an emulator cannot reproduce completely. Optical noise, calibration, component tolerances, numerical precision, conversion between electronic and optical domains, data movement into and out of the accelerator, and software integration can all affect the performance available from a deployed system.
The structure of the numerical problem also remains important. Sparse matrices can differ substantially in conditioning, distribution of non-zero elements, convergence behaviour, and precision requirements. An architecture that performs particularly well on one class of problem may see a much smaller advantage when those characteristics change.
The study therefore provides a workload-specific indication of where photonic acceleration may offer large gains rather than evidence of uniform acceleration across HPC. Demonstrating the same behaviour in physical hardware across representative applications will determine how closely the projected results translate into deployed systems.
LightSolver’s wider platform presents the LPU through a Python-based interface, with software translating mathematical problems into operations executed by the optical resonator. That abstraction will be important if specialist optical hardware is to be incorporated into computing environments without requiring users to work directly with the underlying photonic controls.
The research gives the architecture a defined numerical target against which physical implementations can be measured. Future hardware results will have to establish not only raw solver speed, but also accuracy, repeatability, integration overhead, and the proportion of end-to-end application time that remains once electronic data preparation and system control are included.
