Hypermultiplexed Integrated Tensor Optical Processor

Optical processors hold great potential to accelerate deep learning tasks with their high clock-rates and low-loss data transmission. However, existing integrated systems are hindered by low scalability due to the quadratic scaling of device counts, energy costs with high-speed analog-to-digital converters, and lack of inline nonlinearity. Here, we overcome these challenges with a wavelength-space-time multiplexed optical tensor processor. Hyperdimensional parallelism allows matrix-matrix multiplications (N^3 operations) using O(N) devices. We incorporated wavelength-multiplexed III/V-based micron-scale lasers (spanning  1 THz) for input activation with inline rectifier (ReLU) nonlinearities and thin-film Lithium-Niobate electro-optic modulators (V_π≈1.3 V) for dynamic weighting. With each device encoding 10-billion activations per second, we demonstrated a machine-learning model with 405,000 parameters. High-clock-rate (10 GS/s), low-energy (500 fJ/OP) parallel computing with real-time programmability unlocks the full potential of light for next-generation scalable AI accelerators.