silicOn photoniCs sTochastic mAtrix multiplier with phase chaNge matErials – OCTANE

This research project will pave the way to next generation computing architectures relying on hardware accelerators based on silicon photonics (SiPh), phase change materials (PCM) and stochastic computing (SC). Indeed, SiPh, reinforced by PCM for non-volatile functions is key to produce low cost and energy-efficient complex and highly integrated photonic devices while SC is perceived as a promising route to a reliable, also low-cost, and low-power alternative to conventional computing approaches.

This project will start by performing two parallel studies, one theoretical and the other experimental, on SC circuits and systems based on PCM integrated on a SiPh platform. The theoretical studies, based on the current PCM state of the art, will propose an innovative system architecture as well as specify the individual underlying photonic devices (with PCM or standard components). The experimental studies will focus on material characteristics for PCM as well as phase transition mechanisms of (i) thin films of material then (ii) PCM patches integrated on silicon photonics waveguides. OCTANE will propose a complete system architecture to demonstrate the applicability of the concept for SC. This circuit, as well as specific test cells (such as individual multipliers with accumulation – MAC), will then be manufactured and experimentally tested. Following this first testing campaign, and based upon the obtained results, an optimized architecture will be built for the individual MAC and SC circuit demonstrator and fabricated. The experimental demonstration of this optimized circuit will constitute the main concluding outcome of OCTANE.

In order to maximize the impact of our results on the scientific community and to project the results to a higher level of complexity than is experimentally feasible within the project, a detailed benchmarking of OCTANE’s results (demonstrated experimentaly or modelled) with conventional or SC based hardware accelerators for neuromorphic computing will also be carried out in a simulation-based approach. Energy gains (energy per MAC, expected to be 10x better than state of the art) for several key applications shall be quantified.