Analog optical computer for AI inference and combinatorial optimization

Kirill P. Kalinin, Jannes Gladrow, Jiaqi Chu, James H. Clegg, Daniel Cletheroe, Douglas J. Kelly, Babak Rahmani, Grace Brennan, Burcu Canakci, Fabian Falck, Michael Hansen, Jim Kleewein, Heiner Kremer, Greg O'Shea, Lucinda Pickup, Saravan Rajmohan, Antony I. T. Rowstron, Victor Rühle, Lee Braine, Shrirang Khedekar, Natalia G. Berloff, Christos Gkantsidis, Francesca Parmigiani, Hitesh BallaniNat 2025 · DOI

TL;DR

Researchers have developed an analog optical computer (AOC) that can perform both AI inference and combinatorial optimization tasks on a single platform. This novel approach leverages a fixed-point search strategy, bypassing energy-intensive digital conversions and demonstrating significant potential for faster and more sustainable computing.

What Problem Does This Paper Address?

The ever-increasing computational demands of Artificial Intelligence (AI) and combinatorial optimization are straining the capabilities and energy efficiency of traditional digital computing. While specialized hardware exists for each domain, few systems can efficiently handle both. Furthermore, many unconventional computing approaches rely on frequent, power-hungry digital-to-analog conversions, and often suffer from mismatches between application requirements and hardware capabilities. This paper tackles these challenges by proposing a unified analog optical platform that can accelerate both AI inference and optimization tasks, aiming for higher efficiency and speed.

Methods

The core innovation is an Analog Optical Computer (AOC) that combines analog electronics and 3D optics. It operates on a fixed-point search principle, iteratively refining a solution until it converges. This process involves:

  • Optical Matrix-Vector Multiplication: This is handled by a microLED array (representing the input vector) and a spatial light modulator (SLM) (representing the weights or problem coefficients). Light from the microLEDs is modulated by the SLM, and the results are summed by a photodetector array. The use of 3D optics allows for inherently parallel and scalable multiplication.
  • Analog Electronic Nonlinearity, Subtraction, and Annealing: After the optical computation, the signal is processed by analog electronics. This includes a hyperbolic tangent (tanh) function for nonlinearity, summing/differencing amplifiers for addition/subtraction, and variable gain amplifiers for annealing (controlling state magnitude reduction) and momentum.
  • Fixed-Point Abstraction: The entire process is framed as finding a fixed point of an iterative update rule. This abstraction is crucial for analog hardware as it is compute-bound, requires minimal intermediate memory, and is inherently more robust to analog noise due to the "attractor" nature of the fixed point.
  • Co-design of Hardware and Abstraction: The paper emphasizes the importance of designing the hardware and the computational abstraction together, mirroring the co-evolution of digital accelerators and deep learning models.

Datasets

The paper utilizes standard datasets for its demonstrations:

  • MNIST and Fashion-MNIST: Used for image classification tasks.
  • Synthetic QUMO and QUBO problems: Used for benchmarking optimization capabilities.
  • FastMRI dataset: Used for medical image reconstruction.
  • Industrially relevant transaction settlement scenarios: Generated for financial optimization tasks.

The authors also developed a digital twin (AOC-DT) that achieves over 99% correspondence with the physical hardware, used for training and validation.

Key Insights and Ideas

  • Unified Platform for AI and Optimization: The AOC demonstrates that a single analog optical architecture can effectively accelerate both AI inference and combinatorial optimization, addressing a significant gap in current hardware development.
  • Fixed-Point Search for Noise Robustness: Framing computations as fixed-point searches inherently leverages the analog nature of the hardware, making it more robust to noise compared to other analog approaches.
  • Eliminating Digital-Analog Conversions: By performing computations entirely in the analog domain, the AOC bypasses energy-intensive digital-analog conversions, leading to substantial efficiency gains.
  • Co-design is Key: The success of the AOC highlights the critical role of co-designing hardware architectures with computational abstractions and algorithms for optimal performance.
  • Potential for Sustainable Computing: The projected efficiency improvements (up to 100-fold) position the AOC as a promising path towards more sustainable computing solutions.

Key Results

  • AI Inference:
    • Achieved high accuracy on MNIST and Fashion-MNIST classification tasks, with results closely matching the AOC-DT.
  • Demonstrated the ability to run nonlinear regression tasks, showing good correspondence with ground truth, though with some sensitivity to noise for complex curves.
  • Successfully implemented deep equilibrium models, which are compute-bound on digital hardware but naturally suited for the AOC.
  • Combinatorial Optimization:
    • Solved QUMO problems with up to 64 variables for medical image reconstruction and financial transaction settlement.
  • Achieved state-of-the-art results on some instances from a standard quadratic optimization benchmark, outperforming commercial solvers in speed.
  • Demonstrated a 100% success rate on smaller QUMO instances derived from real settlement data, outperforming quantum hardware in some cases.
  • Performance:
    • Projected performance of 500 tera-operations per second (TOPS) per watt at 8-bit precision, over 100 times more efficient than leading GPUs.
  • Fixed-point iterations complete in approximately 20 ns.

Limitations and Future Work

  • Current Scale: The current hardware is small-scale, supporting up to 4,096 weights. Scaling to billions of weights for real-world applications will require modular architectures and further miniaturization.
  • Noise Sensitivity in Regression: While generally noise-robust, regression tasks showed slightly more sensitivity to noise, requiring averaging over multiple runs for smooth curves.
  • Training on Digital Twin: Model training is performed digitally using the AOC-DT, with only inference happening on the analog hardware.
  • Future Work: The authors envision scaling the AOC through modularity, tighter integration of optics and electronics, and exploring more complex compute primitives. They also suggest that the AOC's co-design approach can drive future innovations in both hardware and algorithms for AI and optimization.

Summary updated

BibTeX
@article{kalinin2025analog,
  author = {Kirill P. Kalinin and Jannes Gladrow and Jiaqi Chu and James H. Clegg and Daniel Cletheroe and Douglas J. Kelly and Babak Rahmani and Grace Brennan and Burcu Canakci and Fabian Falck and Michael Hansen and Jim Kleewein and Heiner Kremer and Greg O'Shea and Lucinda Pickup and Saravan Rajmohan and Antony I. T. Rowstron and Victor Rühle and Lee Braine and Shrirang Khedekar and Natalia G. Berloff and Christos Gkantsidis and Francesca Parmigiani and Hitesh Ballani},
  title = {{Analog optical computer for AI inference and combinatorial optimization}},
  journal = {Nat.},
  year = {2025},
  doi = {10.1038/s41586-025-09430-z},
}
Figure