Design of Flexible Domain-Specific Accelerators for Diverse and Dynamic Computation

In today’s era of computing, the massive amounts of data generated and processed, along with advanced algorithms, have driven the development of hardware accelerators. These accelerators offer high performance and energy efficiency, outperforming CPUs and GPUs. Over the past decade, Application-Specific Integrated Circuits (ASICs) have been widely developed for deep learning.

However, the progress in algorithms and growing application demands have increased heterogeneity and complexity in modern computing. These new workloads require various computation kernels, rely on dynamic decision-making, and change kernel usage at runtime. Additionally, the shift to edge computing, with stricter energy and area constraints, further complicates accelerator designs that can support such diverse and dynamic workloads.

Domain-specific accelerators have emerged as a versatile solution. They provide the programmability and flexibility needed to address the growing challenges while maintaining performance and energy efficiency. This dissertation presents three domain-specific accelerator prototypes for diverse and dynamic computation: VOTA, a programmable accelerator that supports heterogeneity in visual object tracking; eNODE, a flexible accelerator to address the computational complexity of dynamic system modeling; and EVA, an evolvable accelerator that supports a variety of workloads in edge computing with runtime adaptation.

CHAIR: Professor Zhengya Zhang