Artificial intelligence (AI) is humanity's new frontier. The state-of-the-art approach to AI relies on artificial neural networks, which learns from vast amounts of data and have achieved tremendously success in tasks ranging from identifying photos to perceiving the environment of a moving car. Today, most AI tasks are done electronically, using electronics integrated circuits, which are very advanced but are reaching limits in speed, energy consumption, and heat dissipation. The age of AI calls for new computing architectures, devices, and integration technologies for efficient computing.
The DOLORES project aims to explore and develop a new computing approach in which light carries and processes the information on a photonic chip for the computing, targeting high compute throughput and energy efficiency. Instead of running these calculations only in electronic circuits, DOLORES uses a photonic chip and is designed to combine the best of photonics and electronics to perform key mathematical operations, such as matrix multiplication, which is one of the most important operations in AI. Many existing optical neural network processors use analog signals and architectures, which can be fast but often suffer from limited numerical precision, degrading the neural network performance. DOLORES takes a different approach, and the main objective is to develop a new type of digital optical computing platform for neural networks for high-precision, reliable optical computing and compatibility with modern electronics. In simple terms, the project tries to answer the question: Can we build an optical AI processor that is both efficient and accurate to be practical enough for real-world applications? To achieve this, DOLORES combines several technologies supported by partners within the consortium, including (1) PIC design, fabrication, and post-fabrication trimming; (2) highly energy-efficient multi-wavelength light source; (3) photonic-electronic co-optimised analog-to-digital converters; (4) advanced chip packaging assembly; and (5) neural network compression.
Integrating the best of photonics and electronics, with expertise brought together in the consortium, DOLORES is expected to advance optical computing toward practical AI hardware, and in the long-term vision enable new possibilities that are beyond the practical limits of today’s computing hardware. If successful, DOLORES can pave the way for a new era in AI, drastically accelerating neural networks and deep learning applications, which could eventually benefits sectors of our society, including healthcare, finance, transportation, and communications.
