Using Interferometer-Based Photonic Integrated Circuits to Perform Rapid Unitary Transformations

Artificial intelligence (AI) has been growing exponentially over the past two decades. AI-based computational demands have thus been skyrocketing, while at the same time, Moore’s Law is beginning to end. Thus, we introduce a transistor-independent optical chip able to perform matrix-vector multiplication. Matrix-vector multiplication is fundamental for computation within AI, graphics, machine learning, image processing, computer vision, graph theory, and quantum computing. We specifically discuss a layout capable of performing a unitary matrix transform by routing light through a lattice of Mach–Zehnder interferometers. We discuss how this transformation can be built up to any arbitrary matrix-vector multiplication. For a given 4 \(\,\times \,\) 4 unitary matrix, the chip performs at 7 teraflops with a 245 teraflops/watt performance per watt rating. Data analysis from simulations and testing is shown, including measured loss at different wavelengths.