Xu, Y. (CSE) – Right Place, Right Time: Accelerating Edge Computation on Modern Heterogeneous SoCs

Modern edge computing increasingly relies on heterogeneous System-on-Chip (SoC) architectures. These chips tightly integrate general-purpose CPUs with various specialized accelerators, including GPUs, FPGAs, and AI accelerators, all under a shared memory architecture. Although these shared-memory SoCs enable more efficient communication and data sharing between different processing units, they are notoriously difficult to program and tune due to architectural diversity across vendors and asymmetric compute capabilities within each SoC.

This dissertation introduces Redwood and BetterTogether, two frameworks that rethink CPU-accelerator collaboration on heterogeneous SoCs. Redwood targets a class of algorithms termed traverse–compute, that combine irregular tree traversals with dense leaf-level computation, e.g., Nearest-Neighbor Search and Barnes–Hut algorithm.

It addresses the efficient mapping of these algorithms onto heterogeneous systems by exploiting the architectural strengths of CPUs, GPUs, and FPGAs. BetterTogether extends this methodology to a different class of edge workloads, specifically multi-stage pipelines and neural networks commonly used in computer vision tasks. Furthermore, it introduces interference-aware analysis and scheduling techniques tailored for mobile SoCs. Finally, to broaden the scope of heterogeneous acceleration, we evaluated emerging domain-specific accelerators. We provide a preliminary analysis of Tensor Processing Units and Tensor Cores within the context of modern programming abstractions.

Event Host: Yanwen Xu, Ph.D. Candidate, Computer Science and Engineering

Advisor: Tyler Sorensen

Zoom- https://ucsc.zoom.us/j/5354629158?pwd=0CVhbwLuXDMX5fAGZd63tcfNqDWp0t.1

Passcode- 114514