Author ORCID Identifier

https://orcid.org/0000-0003-1847-3976

Date Available

5-10-2024

Year of Publication

2024

Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation

College

Engineering

Department/School/Program

Electrical Engineering

First Advisor

Ishan G Thakkar

Abstract

Artificial Intelligence (AI) has experienced remarkable success in recent years, solving complex computational problems across various domains, including computer vision, natural language processing, and pattern recognition. Much of this success can be attributed to the advancements in deep learning algorithms and models, particularly Artificial Neural Networks (ANNs). In recent times, deep ANNs have achieved unprecedented levels of accuracy, surpassing human capabilities in some cases. However, these deep ANN models come at a significant computational cost, with billions to trillions of parameters. Recent trends indicate that the number of parameters per ANN model will continue to grow exponentially in the foreseeable future. To meet the escalating computational demands of ANN models, the hardware accelerators used for processing ANNs must offer lower latency and higher energy efficiency. Unfortunately, traditional electronic implementations of ANN hardware accelerators, including CPUs, Graphics Processing Units (GPUs), Application-Specific Integrated Circuits (ASICs), and Field Programmable Gate Arrays (FPGAs), have fallen short of meeting the latency and energy efficiency requirements for processing deep ANN models. Furthermore, the interconnection network subsystems in these electronic accelerator systems, designed to facilitate large-scale data transfers between processing cores and memory/control units within the accelerator systems, have become bottlenecks that hinder the throughput, latency, and energy efficiency of deep ANN model processing.

Fortunately, Photonic Integrated Circuits (PICs)-based accelerator systems, featuring photonic network subsystems are promising alternatives to conventional electronic accelerators. PIC-based accelerator systems operate in the optical domain, delivering processing at the speed of light with ultra-low latency, minimal dynamic energy consumption, and high throughput. These advantages stem from the wavelength division multiplexing capabilities and the absence of distance-dependent impedance in PICs. Furthermore, these characteristics enable the implementation of high-performance photonic network subsystems within PIC-based accelerator systems. Additionally, PIC-based accelerator systems offer inherent optical nonlinearities.

Despite these numerous advantages over electronic accelerators, PIC-based systems still encounter several challenges due to limited optical power budget, susceptibility to crosstalk and other sources of noise caused by the analog operation, high area consumption, and restricted functional flexibility of PICs. These challenges manifest in various ways. (i) The existence of a significant trade-off between the achievable processing core size and the supported bit precision that impedes the scalability of processing cores. (ii) The limited reconfigurability, in terms of supported computing size and precision, makes them less adaptable to modern ANN models with diverse computational and precision demands. (iii) The reliance on electronic adder networks for accumulation diminishes the latency and energy consumption benefits of PIC-based accelerator systems due to frequent analog-to-digital conversions and memory accesses involved in accumulations.

My research has contributed several solutions that overcome a multitude of these challenges and improve the throughput, energy efficiency, and flexibility of PIC-based AI accelerator systems. I identified and analyzed factors that affect the scalability and reconfigurability of PIC-based AI accelerator systems. I proposed several novel PIC-based accelerator architectures with enhancements at the circuit level, architecture level, and system level to improve scalability, reconfigurability, and functional flexibility. At the circuit level, these enhancements serve to decrease optical signal losses, reduce control complexity, enable adaptability for various ANN processing tasks, and lower power and area consumption. The architecture-level improvements mitigate crosstalk noise, facilitate functional reconfigurability, enable in-situ and flexible spatio-temporal accumulation, and provide flexible support for different dataflows. The system-level enhancements involve the integration of stochastic computing with PIC-based accelerators to break the inherent trade-off between scalability and supported bit precision. Additionally, applying stochastic computing enhances the flexibility of PIC-based accelerators, allowing them to support mixed-precision ANN models. These cross-layer enhancements collectively contribute to the design of PIC-based AI accelerator systems, resulting in improved throughput, energy efficiency, scalability, and reconfigurability.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2024.143

Funding Information

My research was supported by the National Science Foundation (NSF), CNS-2139167

Download

Share

COinS