Streamlining High-Performance Heterogeneous Hardware Design

Hybrid Event: 3725 BBB /  Zoom  Passcode: goblue

Abstract: With the stagnation of Moore’s Law and the breakdown of Dennard Scaling, hardware designers are increasingly adopting heterogeneous architectures to achieve superior performance and energy efficiency. These architectures integrate specialized hardware components tailored to specific applications, significantly outperforming traditional homogeneous designs. Such specialized components have been instrumental in recent technological advances, notably enabling real-time deep-learning inference by reducing latency from hours to seconds.

However, the advantages of heterogeneous systems come with substantial costs, including increased non-recurring engineering efforts, high communication overhead, and memory bandwidth contention—issues that limit performance and hinder wider adoption. A significant contributor to these issues is inefficient communication between heterogeneous components. Specifically, mismatches in I/O interfaces reduce reusability and elevate engineering costs; inadequate tolerance to communication latency introduces critical performance bottlenecks; and ineffective allocation of communication resources exacerbates contention and interference.

This dissertation presents novel, minimally invasive approaches to streamline the design of high-performance heterogeneous hardware by directly addressing these communication-related challenges. First, it introduces automated methods for generating heterogeneous hardware designs using standardized communication interfaces, greatly enhancing design reusability and speeding up iterative development. Second, it proposes latency-tolerant bus optimization techniques, allowing systems to tolerate microsecond-level delays without intrusive redesign. Finally, it presents a flexible interconnect design with distributed resource allocation, efficiently managing communication resources to improve system performance and reduce contention.

By directly confronting these key obstacles, the approaches developed in this dissertation enhance the practicality, efficiency, and overall performance of heterogeneous hardware designs, thus laying a foundation for future advancements and innovations in the field of heterogeneous architecture.