Leveraging million-scale Non von Neumann computations for accelerated Machine Learning and High Performance Computing

Abstract

Current large-scale computations, for instance in High Performance Computing or in the training of massive Machine Learning models, often suffer from the “memory bottleneck”, especially when dealing with high-dimensional data. Here, we present a new non-von Neumann photonic hardware, leveraging multiple light scattering. Optical Processing Units can be seamlessly integrated into a variety of hybrid photonics / silicon pipelines implementing state-of-the-art Machine Learning or High Performance Computing algorithms. They offer a credible pathway towards a new generation of large-scale computing, both scalable and sustainable.