UC Santa Barbara

The explosive growth of data volume brought by the pervasiveness of artificial intelligence (AI) applications is calling for interconnect technologies that enable higher bandwidth capacity at a lower cost. In particular, optical interconnects based on silicon photonics are considered a promising substitute for electrical ones, given their cost-effectiveness and scalability enabled by a CMOS-compatible fabrication process. However, as the optical interconnects further penetrate the shorter-reach regime, several issues arise from the growing complexity of system integration and pose challenges to their design quality, including 1) inadequate support from design automation methodologies for the modeling and simulation of the optical interconnects, 2) oversimplified characterization of process variations, resulting in variation alleviation techniques with limited effectiveness, and 3) the lack of runtime reconfiguration strategies for the optical interconnects under traffic dynamics, leading to unoptimized energy efficiency. This dissertation is devoted to addressing the above issues by solutions proposed at the device, link, and system levels, paving the way to the quality design of variation-aware runtime-reconfigurable optical interconnects with optimized energy efficiency.

The first part of this dissertation focuses on device-level methodologies for electronic-photonic design automation (EPDA), including compact models developed for lasers and modulators and a novel hierarchical spatial variation model characterized for silicon microring resonators. Extensively validated by measurement data, the library of device-level models enables accurate circuit-level simulation of optical links and variation-aware estimation of the link power budget, serving as the fundamentals of the optimization techniques proposed at the link and system levels.

The second part of this dissertation proposes three link-level techniques to improve the energy efficiency of the optical interconnects under wafer-scale process variations. The three techniques exploit, respectively, 1) sub-channel redundancy of carrier wavelengths, 2) a combination of electrical and thermal tuning mechanisms, and 3) optimal mixing and matching of a batch of fabricated transceivers, achieving significant reductions in the energy required for transmitting a single bit of data.

The third and final part of this dissertation proposes two strategies at the system level that further improve the energy efficiency of the optical interconnects under traffic dynamics by reconfiguring the link power at application runtime. The two strategies incorporate assistance from 1) traffic adjustment enabled by task mapping exploration and 2) traffic adaptability enabled by traffic prediction, respectively, and achieve substantial improvements in energy consumption with minimal overhead to application execution time, notably outperforming existing strategies.