With an ongoing trend in computing hardware toward increased heterogeneity, domain-specific coprocessors are emerging as alternatives to centralized paradigms. Here I will introduce and discuss key concepts on PIC-based photonic paradigms for machine intelligence. These include parallelization strategies, real-time computing (~ps), non-iterative one-shot execution O(1) run-time complexities, synergistic convolutions, photonic nonvolatile (multistate) memory. These device, circuit, and system design paradigms allow realizing PIC-based special-purpose processors in this ongoing trend towards increased heterogeneity. As an example, I will share a photonic tensor core processor and a massively-parallel convolutional neural network processor.
|