Direct tensor processing with coherent light (www.nature.com)

Researchers introduced POMMM (parallel optical matrix–matrix multiplication), a new paradigm that performs full tensor processing in a single coherent-light propagation. POMMM encodes matrix rows as spatial amplitude patterns with distinct linear phase gradients, applies a column-wise Fourier transform to create a “hybrid” superposition, modulates amplitude with the transposed second matrix, and finishes with a row-wise Fourier transform so different row contributions separate into distinct spatial-frequency positions. A benchtop prototype using amplitude and phase spatial light modulators, cylindrical-lens assemblies and a qCMOS sensor validated the approach, producing single-shot outputs for entire matrix products without time/space/wavelength multiplexing. This method is significant because it scales matrix–matrix multiplication performance with data dimension while exploiting optics’ innate bandwidth, parallelism and energy efficiency—addressing GPU bottlenecks like memory bandwidth and power for ML workloads. Experiments and simulations show strong agreement with GPU MMM across real and complex matrices (mean absolute error <0.15, NRMSE <0.1) and successful direct optical inference for CNN and ViT components (multi-channel convolution, multi-head attention, multi-sample FC) on MNIST/Fashion-MNIST. POMMM supports multi-wavelength expansion, has analyzed error sources with mitigation strategies, and—by avoiding repeated propagations—promises higher theoretical efficiency and generality than prior optical computing schemes, positioning it as a scalable foundation for next-generation optical neural-network acceleration.