HAETAE will a) deploy a multi-material PIC platform that combines the best-in-class computational technologies, using Si3N4 as its low-loss interconnection layer, Si/SiGe as its high-speed computational layer, InP actives for on-chip gain and nonlinearities, and non-volatile Micro-Electro-Mechanical Systems (MEMS) for zero-energy static weighting. This unique synergy will be facilitated through a dual-folded heterogenous integration approach, that synergizes micro–Transfer Printing (uTP) of InP semiconductor optical amplifiers (SOAs) on a 3D Si/ Si3N4 platform and MEMS-based cross-platform compatibility through multi-chiplet bonding co-integration. b) align this along the lowest-loss 3D Si/Si3N4 WDM-equipped coherent optical Crossbar (Xbar) architecture further enhancing the existing Xbar for highly scalable MVM chiplets, c) develop InP all-optical reconfigurable Non-Linear Activation Function (NLAF) circuits for connecting cascaded photonic neural layers without requiring energy-hungry digital interfaces or O/E/O inter-layer conversions while also providing on-chip optical-gain d) align its development roadmap along the primitives of the Transfer Learning (TL) paradigm that is widely used for optimizing and adapting large AI models like ChatGPT and Meta’s MultiRay to various tasks. It typically involves using a static pre-trained backbone neural network (NN) followed by shallow retrainable NN layers, e.g., Large Language Models (LLM)-oriented adapters. This forms an ideal DL framework for converging AI with energy-efficient photonic neuromorphic hardware, mainly utilizing low-energy statically-weighted photonic neurons in its pre-trained backbone segment yet supporting a variety of applications through a limited number of photonic Tiled Matrix Multiplication (TMM) engines in its retrainable layer part. HAETAE demonstrators will exhibit significant improvements across all critical metrics of state-of-the-art AI chipsets: energy efficiency, requiring only 18.5 fJ/MAC, ii) area efficiency, offering 2.45 TMAC/s/mm2, and iii) computational power, unleashing an impressive ~4.1 PMAC/s! HAETAE will highlight the wide application perspectives of its photonic TL (PTL) setups by validating them in three challenging use-cases in the areas of (i) DL and AI computing as a LLM transformer, (ii) Data Center (DC) cybersecurity threat detection applications, and (iii) Photonic AI-assisted DSP for fiber communications.
