Artificial Intelligence at the Speed of Light

The massive data deluge from mobile, IoT, and edge devices, together with powerful innovations in data science and hardware processing, have established artificial intelligence (AI) as the cornerstone of modern medical, automotive, industrial automation, and consumer electronics domains. Domain-specific AI accelerators such as Google’s TPU, Apple’s Bionic, and Intel’s Nirvana, now dominate CPUs and GPUs for energy-efficient AI and machine learning processing. However, the evolution of these electronic accelerators is facing fundamental limits due to the slowdown of Moore’s law and the reliance on metal wires, which already severely bottleneck computational performance today. Silicon photonics represents a promising post-Moore technological alternative to overcome these limitations. Not only can photonic interconnects fabricated in CMOS-compatible processes provide near speed of light transfers at the chip-scale, but photonic devices can now also perform computations entirely in the optical domain. In this talk, I will present my vision of how silicon photonics can drive an entirely new class of sustainable AI hardware accelerators that can provide orders of magnitude energy improvements over today’s accelerators. I will discuss the evolution of silicon photonics over the past two decades, from integrated optics to photonic devices that can now be fabricated with low-cost CMOS-compatible manufacturing techniques. I will then cover new directions in power minimization, variation tolerance, fault resilience, and security for communication and computation with silicon photonics. I will share experiences from my journey over the past decade and a half towards the goal of realizing viable silicon photonic networks and computing substrates. I will end the talk with a discussion of the EDA related and other challenges and opportunities to achieve unparalleled energy-efficiency and performance gains in future
manycore computing platforms with silicon photonics.