• About
    • Our Team
    • Careers
    • Stories
  • Technology Transfer
    • Available Technologies
    • Innovators
      • Meet Our Innovators
      • Submit Disclosure
      • NAI Chapter
    • Startups
      • Meet Our Startups
    • Lab to Life
      • About Lab to Life
      • Lab to Life Process
      • Lab to Life Startups
      • Lab to Life Team
      • Lab to Life Contact
    • AI POC Grant Program
    • Workshops and Events
      • CSU Demo Day
    • FAQ
  • Real Estate Services
    • The Prospect
    • Maintenance Request Form
    • Commercial Leasing
  • Financing Program
  • Maxwell Ranch
  • Connect
  • Newsletter Signup
  • Submit Disclosure
  • About
    • Our Team
    • Careers
    • Stories
  • Technology Transfer
    • Available Technologies
    • Innovators
      • Meet Our Innovators
      • Submit Disclosure
      • NAI Chapter
    • Startups
      • Meet Our Startups
    • Lab to Life
      • About Lab to Life
      • Lab to Life Process
      • Lab to Life Startups
      • Lab to Life Team
      • Lab to Life Contact
    • AI POC Grant Program
    • Workshops and Events
      • CSU Demo Day
    • FAQ
  • Real Estate Services
    • The Prospect
    • Maintenance Request Form
    • Commercial Leasing
  • Financing Program
  • Maxwell Ranch
  • Connect
  • Newsletter Signup
  • Submit Disclosure

LightPro: The Ultra-Compact, Low-Power AI Photonic Processor

At a Glance

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. 

Background

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. 

Overview

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. 

Three bar charts compare LightPro and MZI-based Clements photonic networks at different matrix sizes (N = 4, 8, 16, 32). The first chart shows LightPro has substantially smaller footprint area than the Clements network as size increases. The second chart shows lower active power consumption for LightPro and pruned LightPro compared to Clements, with power rising with matrix size. The third chart shows similar high accuracy for both networks across sizes and on MNIST.
Figure 1. (a) The Footprint area for the LightPro network compared to its equivalent MZI-based Clements network. (b) The total active power consumption related to phase shifters for an optimized and pruned LightPro network compared to the MZI-based Clements network with different sizes when the same multiplication was implemented. (c) Accuracy results of a pruned LightPro network compared with the MZI-based Clements network.

Benefits

  • Nonvolatile Operation: Uses PCMs to maintain programmed states without continuous power, drastically reducing static energy consumption.
  • Massive Footprint Reduction: Achieves up to an 85% smaller area compared to traditional MZI-based meshes, enabling higher integration density.
  • High Scalability: Minimizes accumulated optical losses and phase noise by using more compact, low-loss directional couplers.
  • Optimized Architecture: Employs NAS and pruning to eliminate redundant components, further enhancing efficiency and robustness against fabrication variations.
  • Accuracy: Experimental verification shows strong agreement with simulations, maintaining nearly identical accuracy to original trained weights

Applications

  • AI Hardware Accelerators: Energy-efficient MVM engines for next-generation deep neural network training and inference.
  • Edge Computing: Compact, low-power processors for AI tasks in autonomous systems, mobile devices, and IoT sensors.
  • Data Center Interconnects: High-speed, high-bandwidth optical communication and signal processing.
  • Neuromorphic Computing: Hardware realization of brain-inspired optical neural networks.
  • Real-time Diagnostics: Fast processing for medical imaging, early cancer diagnosis, and pandemic forecasting.

Publications

A. Shafiee, et al. (2025) “LuxNAS: A Coherent Photonic Neural Network Powered by Neural Architecture Search.” Optica Publishing Group. https://doi.org/10.1364/CLEO_AT.2025.JPS100_93  

A. Shafiee, et al. (2025) “LightPro: A Linear Photonic Processor with Full Programmability.” https://doi.org/10.21203/rs.3.rs-7862546/v1

 

Last Updated: February 2026
A high-tech, stylized icon of a square photonic computer chip on a dark background. Inside the chip, bright blue lines of light intersect in circular and elliptical orbits, connected by glowing red and orange nodes that symbolize thermal control and data processing.
Opportunity

Available for Exclusive Licensing  
TRL: 4

IP Status

Nothing filed yet

Inventors

Mahdi Nikdast
Amin Shafiee 

Reference Number
2025-092
Licensing Manager

Jessy McGowan
Jessy.McGowan@colostate.edu
970-491-7100

Contact Us About this Technology
Download PDF
Download
Strata logo white
  • Technology Transfer
  • Real Estate Services
  • Financing Program
  • Maxwell Ranch
  • Connect
  • Equal Opportunity Employer
  • Tax Information
  • Technology Transfer
  • Real Estate Services
  • Financing Program
  • Maxwell Ranch
  • Connect
  • Equal Opportunity Employer
  • Tax Information
Serving the Colorado State University System with strategic real estate services, intellectual property protection and licensing, and financing activities.
Integrity
Reliability
Respect
Innovation
Excellence
Submit Disclosure

Newsletter Sign Up

Connect