Dissertation Title: Behavior Specialized Processors
Karthikeyan Sankaralingam (Advisor)
Improvements in the performance and energy consumption of general purpose processors have slowed dramatically over the last decade. This is due to the combined effect of breakdowns in transistor scaling, causing severe chip-level power limitations, and monolithic and inefficient general purpose microarchitecture. Over the last decades, and especially in recent years, the community has turned towards domain specific processors, where general purpose programmability is jettisoned for efficiency. These trends threaten the future of general purpose architecture innovation.
This dissertation explores a promising alternative to the domain-specific approach: to specialize for properties of programs which span across domains, which we refer to as program behaviors. Programmable hardware engines which exploit these characteristics -- behavior specialized accelerators -- can be integrated into general purpose cores, and used as offload engines to transparently improve their performance and energy efficiency during amenable program phases. This work addresses challenges in accelerator modeling, microarchitecture, compilation, and design-space exploration.
I will discuss several key findings. First, that a small number of exploitable program behaviors can be used to characterize a majority of applications. Second, that dataflow architectures become practical and useful in hybrid execution with a general purpose core. Third, synergistic behavior-specialized accelerators combined with simple general core pipelines can facilitate disruptive microprocessor tradeoffs, enabling mobile-class processor energy-efficiency with desktop-class performance. In addition to the architectural discoveries, this dissertation proposes a modeling methodology which enables rapid exploration of behavior specialized processors, as well as a mathematical formulation for declarative and optimal instruction scheduling on these architectures. Overall, the paradigm of behavior specialization demonstrates that the future for general purpose architecture innovation is bright.