Universal Linear Optics for Ultra-Fast Neuromorphic Silicon Photonics Towards fJ/MAC and TMAC/sec/mm² Engines

The field of neuromorphic photonics has been projected to comprise the next-generation Neural Network platform, expected to lead to remarkable advances in compute energy- and area-efficiency metrics. Herein, we review the performance of state-of-the-art neuromorphic photonic demonstrators, summarizing the impact of the circuit architecture and employed weight technology on the system credentials in terms of scalability, energy- and footprint-efficiency. We provide an overview of a recently demonstrated photonic crossbar multi-port interferometer, holding significant insertion loss, technology versatility and robustness advantages over all state-of-the-art linear optical layouts. This novel linear optical circuit architecture is then transferred onto an integrated silicon photonic (SiPho) platform, realizing a single-column crossbar and selecting SiGe electro-absorption modulators (EAM) for both its fan in and weighting stages. This single-neuron coherent SiPho prototype is then experimentally benchmarked on the MNIST dataset, allowing for record-high compute-rates up to 50 GHz/axon. The classification accuracy was studied for compute rates ranging between 16-50 GHz, revealing 99.03% accuracy at 16 GHz that reduces by only 3.79% at 50 GHz. Finally, we investigate the interdependence of compute rate, bit resolution and energy efficiency, in photonic accelerator layouts and benchmark our proposed Xbar architecture against state-of-the-art photonic and electronic accelerators. The analysis reveals an energy- and footprint-efficiency of 54 fJ/MAC and 1.54 TMAC/s/mm2, respectively, shaping in this way a promising roadmap for next-generation neuromorphic accelerators.