Building User-Driven Egde Devices

Abstract

Edge devices like smartphones, wearables, and personal assistants have become an integral part of our daily routines. Their ubiquitous and portable nature allows them to operate in any sort of environment. They can be deployed in the wild or at home without requiring a constant power source plugged into them. However, the small form factor and resource-constrained nature of edge devices limit their computation capabilities and, thus, significantly impact the efficiency of tasks performed on edge devices. The application efficiency is directly related to the quality of user experience for the hand-held edge devices; thus, these shortcomings of the edge device impact the user experience as well.

In this dissertation, we develop solutions to address these limitations of edge devices to enhance the performance and energy efficiency of a wide range of user applications processed on edge devices. Our proposed solutions are either low-cost alternatives that can replace expensive silicon or extract extreme efficiencies from already in-use silicon, thus lowering the total cost of ownership of edge devices. Our solutions are driven by three key strategies: 1) cross-component optimizations across the system, 2) leverage user information and preferences in the hardware, and 3) co-design the application and hardware for the edge system. In our first solution, Seesaw, we study user applications for edge devices with tiny microcontrollers and sensors. We propose an end-to-end automated technique to find optimal compute/sensing rates for power-intensive sensors governed by low-power sensors and based on individual users' preferences and inherent sensing capabilities. This elongates battery life with minimal impact on the perceivable user experience.

In our second proposed solution, we customize the machine learning-based image recognition application for each user by creating small and accurate user-specific machine learning models on the resource constraint edge device. This significantly lowers computation demands and memory footprint without impacting user accuracy. %We offer an end-to-end system that identifies user preferences and builds a user-specific customized model to enhance the efficiency of the application on the user's edge device. In the next work, Duet, we leverage the user history and profile information to decompose the giant monolithic recommendation model into a separate user and item model. The user model processes user information in a lightweight manner on the local edge device, and its computation is reused by the item model, processing 100s of items at the datacenter. Thus, we offer enhanced privacy along with performance improvement of 6.4x and energy efficiency of 4.6x.

Finally, we present a low-cost and heterogeneous System-in-Package (SiP)-based multi-chiplet interconnect architecture built over the 2.5D stacking interposer technology, which can replace the expensive monolithic system-on-chip (SoC). The proposed architecture exposes high-bandwidth links of the interposer over which we efficiently map popular bandwidth-intensive edge applications to enhance performance and energy efficiency.