FirePower: Towards a Foundation with Generalizable Knowledge for Architecture-Level Power Modeling

TL;DR

This work proposes a new power modeling solution FirePower that targets few-shot learning scenario for new target architectures and proposes multiple new policies to utilize cross-architecture knowledge.

Abstract

Power efficiency is a critical design objective in modern processor design. A high-fidelity architecture-level power modeling method is greatly needed by CPU architects for guiding early optimizations. However, traditional architecture-level power models can not meet the accuracy requirement, largely due to the discrepancy between the power model and actual design implementation. While some machine learning (ML)-based architecture-level power modeling methods have been proposed in recent years, the data-hungry ML model training process requires sufficient similar known designs, which are unrealistic in many development scenarios. This work proposes a new power modeling solution FirePower that targets few-shot learning scenario for new target architectures. FirePower proposes multiple new policies to utilize cross-architecture knowledge. First, it develops power models at component level, and components are defined in a power-friendly manner. Second, it supports different generalization strategies for models of different components. Third, it formulates generalizable and architecture-specific design knowledge into two separate models. FirePower also supports the evaluation of the generalization quality. In our experiments, FirePower can achieve a low error percentage of 5.8% and a high correlation R of 0.98 on average only using two configurations of target architecture. This is 8.8% lower in error percentage and 0.03 higher in R compared with directly training McPAT-Calib baseline on configurations of target architecture.

Figures (8)

• Figure 1: Proposed power modeling paradigm FirePower vs. existing architecture-specific paradigm. FirePower targets few-shot learning for new target architectures. It extracts general knowledge from an already known architecture, providing a "foundation" to support modeling new architectures.
• Figure 2: The illustration of our power-friendly component definition for the out-of-order CPU core.
• Figure 3: Correlation between power and the most related hardware parameter. The two components correlate with different hardware parameters. DCacheDataArray correlates with $DCacheWay$, DCacheMSHR correlates with $MSHREntry$.
• Figure 4: The FirePower framework with two phases. Knowledge extraction in phase 1 extracts hardware model and parameter importance from a known architecture as general knowledge. Application in phase 2 adopts two knowledge generalization strategies, Retraining and No Retraining, depending on the parameter importance distribution.
• Figure 5: Summary of the comparison between FirePower and other methods under different numbers of configurations of target architecture. "Comp" stands for Component and "Transfer" stands for Transfer Learning.