will develop novel strategies for joining key photonic foundry platforms in hybrid systems.
Ultra low loss SI3N4 integrated photonic circuits for comb generation (EPFL).
Therefore, the PHOENICS project focuses on three main objectives:
1) The PHOENICS consortium will establish photonic neuromorphic processors as a new paradigm for AI computing with unprecedented performance and energy efficiency.
These processors will drive applications where low latency, high bandwidth, and low energies are paramount. Prominent applications include nonlinear programming (e.g. solving optimization problems and partial differential equations), scientific (e.g. protein folding simulations) and high-performance computing (e.g. vector-matrix multiplications), machine learning acceleration (e.g. deep learning inference, and ultrafast and online learning), and intelligent signal processing (e.g. generation of wideband RF signals, fiber transmission equalization, spectral mining).
2) The PHOENICS consortium will explore novel materials to implement neuro-mimicking photonic components and compute principles.
PHOENICS aims to develop disruptive hybrid architectures for photonic processing by joining foundry-ready circuit-level implementations with phase-change materials. Hybrid phase-change photonic devices provide the capability for non-volatile, all-optical data storage, which is paramount for implementing learning systems and inference devices. The use of phase-change materials enables ultra-high switching speeds compatible with photonic processing speeds, paired with high cyclability and endurance. These attributes are key enablers for brain-inspired computing platforms, requiring photonic equivalents of artificial synapses and artificial neurons. Within the PHOENICS project, we will develop key elements of such photonic hardware mimics of neural systems.
3) The PHOENICS consortiums will develop approaches for upscaling computational power using wavelength division multiplexing and foundry processing.
The envisaged approach for PHOENICS will be built on two complementary foundry platforms for active photonics using InP and silicon photonics. The PHOENICS project will generate highly parallel input vectors using integrated frequency microcombs which are modulated at 20 GHz speed to deliver ultra-high data rates to the photonic tensor core. These input signals will be processed in parallel using a silicon photonics matrix processor in which programmable matrix elements are implemented using phase-change materials (acting as artificial synapses).