Researchers at Colorado State University have developed a new type of programmable photonic processor called LightPro that uses light instead of electricity to perform complex calculations. This technology uses special “phase-change materials” that act like a reusable memory, allowing the chip to remember its settings without needing a constant power supply. By organizing these components with a smart search algorithm, the researchers created a system that is much smaller and more energy-efficient than current designs. This innovation could lead to faster, greener AI hardware for everything from smartphones to large data centers.
Matrix-vector multiplication (MVM) is the primary computational bottleneck in deep neural networks (DNNs), accounting for the majority of energy consumption during AI inference and training. While silicon photonics offers a high-speed, energy-efficient alternative to electronic accelerators, traditional coherent designs like Mach-Zehnder Interferometer (MZI) meshes face scalability issues due to their large physical footprint and high active power requirements. LightPro addresses these limitations by integrating nonvolatile phase-change materials (PCMs) into directional couplers, enabling a reconfigurable architecture that maintains its state without continuous power draw.
LightPro is a fully programmable linear photonic processor that leverages Sb2Se3, a low-loss phase-change material, to modulate the coupling coefficients of silicon photonic directional couplers (DCs). By thermally inducing phase transitions between amorphous and crystalline states, the system achieves dynamic control over optical splitting ratios without the constant energy consumption required by traditional thermo-optic phase shifters. To maximize the efficiency of the hardware, the researchers utilized a Neural Architecture Search (NAS) and pruning algorithm that optimizes the network topology by cascading tunable DCs with phase shifters to match specific target unitary matrices.
The technology demonstrates significant quantitative improvements over standard architectures like the Clements mesh. Simulation results indicate that LightPro can achieve up to an 85% reduction in on-chip footprint and more than a 50% improvement in active power consumption. Furthermore, the pruning technique allows for the removal of up to 54% of phase shifters and 55% of DCs for stationary weight matrices, yielding a 67% improvement in power efficiency while maintaining high computational accuracy with less than a 5% drop in performance.
Available for Exclusive Licensing
TRL: 4
Nothing filed yet
Mahdi Nikdast
Amin Shafiee
Jessy McGowan
Jessy.McGowan@colostate.edu
970-491-7100