Learning-based system-level power modeling of hardware IPs

Accurate power models for hardware components at high levels of abstraction are a critical component to enable system-level power analysis and optimization. Virtual platform prototypes are widely utilized to support early system-level design space exploration. There is, however, a lack of accurate and fast power models of hardware components at such high-levels of abstraction.

In this dissertation, we present novel learning‑based approaches for extending fast functional simulation models of white-, gray-, and black-box custom hardware intellectual property components (IPs) with accurate power estimates. Depending on the observability, we extend high-level functional models with the capability to capture data-dependent resource, block, or I/O activity without a significant loss in simulation speed. We further leverage state-of-the-art machine learning techniques to synthesize abstract power models that can predict cycle-, block-, and invocation-level power from low-level hardware implementations, where we introduce novel structural decomposition techniques to reduce model complexities and increase estimation accuracy.

Our white-box approach integrates with existing high-level synthesis (HLS) tools to automatically extract resource mapping information, which is used to trace data-dependent resource-level activity and drive a cycle-accurate online power-performance model during functional simulation. Our gray-box approach supports power estimation at coarser basic block granularity. It uses only limited information about block inputs and outputs to extract light-weight block-level activity from a functional simulation and drive a basic block-level power model that utilizes a control flow decomposition to improve accuracy and speed. It is faster than cycle-level models, while providing a finer granularity than invocation-level models, which allows to further navigate accuracy and speed trade-offs. We finally propose a novel approach for extending behavioral models of black-box hardware IPs with an invocation-level power estimate. Our black-box model only uses input and output history to track data-dependent pipeline behavior, where we introduce a specialized ensemble learning that is composed out of individually selected cycle-by-cycle models with reduced complexity and increased accuracy. The proposed approaches are fully automated by integrating with existing, commercial HLS tools for custom hardware synthesized by HLS. Results of applying our approaches to various industrial‑strength design examples show that our power models can predict cycle‑, basic block-, and invocation-level power consumption to within 10%, 9%, and 3% of a commercial gate-level power estimation tool, respectively, all while running at several order of magnitude faster speeds of 1-10Mcycles/sec.