Analog optical computer for AI inference and combinatorial optimization (www.nature.com)

Researchers have developed an analog optical computer (AOC) that uniquely accelerates both AI inference and combinatorial optimization within a single, energy-efficient platform. Unlike existing systems that either focus on AI or optimization and rely heavily on energy-intensive digital conversions, the AOC operates entirely in the analog domain by combining three-dimensional optics with analog electronics. This approach executes rapid fixed-point iterative searches to perform matrix-vector multiplications optically and nonlinear operations electronically, enabling noise-robust computations without intermediate digital conversions. The unified fixed-point abstraction grounds both neural equilibrium models and quadratic unconstrained mixed optimization (QUMO), addressing memory bottlenecks and application-hardware mismatches prevalent in conventional digital and analog accelerators. The AOC demonstrates its versatility across four diverse use cases—including image classification, nonlinear regression, medical image reconstruction, and financial transaction settlement—using consumer-grade, scalable components. Featured models include emerging equilibrium networks with recursive reasoning capabilities, which are typically compute-bound on digital chips but naturally suited to the AOC's iterative framework. With a prototype supporting up to 4,096 neural weights at 9-bit precision and solving optimization problems involving 64 variables, the system shows state-of-the-art results on benchmark tasks. Projected efficiency exceeds 500 tera-operations per second per watt, making the AOC over 100 times more power-efficient than leading GPUs, signaling a promising direction toward sustainable, scalable analog computing for future AI and optimization challenges.