Hyper‐Parallel Optical Tensor Processors Toward Multitask Neural Networks

ABSTRACT

The exponential growth of multimodal sensor data demands computing architectures capable of massively parallel tensor operations with ultralow latency and high energy efficiency. Optical computing offers a promising route to overcome the von Neumann bottleneck and inherent limits on parallelism in electronic systems. Here, we report a hyper‐parallel distributed multi‐core photonic processor (DMPP) that orchestrates the intrinsic orthogonality of light across spectral and spatial domains to enable multiple‐instruction, multiple‐data processing. Our approach integrates wavelength‐division multiplexing (WDM) with spatially distributed photonic tensor cores, allowing a single processor to process multiple independent data streams and computational tasks concurrently. The system supports multitask workflows by dynamically sharing photonic cores across tasks. This work establishes a scalable photonic computing paradigm for high‐throughput processing of heterogeneous sensor data, enabling efficient, parallel neural network accelerators for edge intelligence and multimodal analytics.