An intra-task dynamic voltage scaling method for SoC design with hierarchical FSM and synchronous dataflow model

Sunghyun Lee, Sungjoo Yoo, Ki‐Young ChoiProceedings of the International Symposium on Low Power Electronics and Design 2003 · DOI

TL;DR

This paper introduces a novel intra-task Dynamic Voltage Scaling (DVS) method for System-on-Chip (SoC) designs that utilize Hierarchical Finite State Machines (HFSM) and Synchronous Dataflow (SDF) models. By precisely identifying execution paths at runtime, their method achieves more optimal energy savings compared to traditional approaches that rely on worst-case assumptions.

What Problem Does This Paper Address?

Designing low-power SoCs is crucial for modern electronics, especially for battery-powered devices. Dynamic Voltage Scaling (DVS) is a popular technique to reduce power consumption by lowering the supply voltage and clock frequency when the system has spare processing capacity. However, achieving optimal DVS requires knowing the exact workload, which is often difficult to determine precisely, especially in complex programs. Existing DVS methods typically assume worst-case or average-case execution paths, leading to suboptimal voltage scaling and thus, less energy savings. This paper tackles the challenge of achieving more precise and efficient DVS in SoCs that employ a combination of hierarchical FSMs for control flow and SDF for dataflow computations.

Methods

The core innovation of this paper lies in exploiting a specific property of the HFSM-SDF model. This model allows for the identification of exact execution paths at runtime before the execution of a schedule for a top-level SDF graph or an FSM state transition.

  1. Runtime Execution Path Identification (RPI): The authors propose an RPI function that, by analyzing the current states of the HFSMs within the model, can predict the precise sequence of SDF actors that will be executed. This is a significant departure from traditional methods that rely on static analysis or worst-case assumptions.
  2. Intra-Task DVS: Once the exact execution path and thus the exact workload are known, a voltage scaler can calculate the optimal supply voltage. This voltage is set at the beginning of each actor's execution to match the required processing power, minimizing energy consumption while still meeting deadlines. The method accounts for the overhead of the RPI and voltage scaling processes.

Datasets

Not discussed in the paper. The authors use a simulated MPEG4 decoder system for their experiments.

Key Insights and Ideas

  • HFSM-SDF's Predictability: The paper highlights that the structure of HFSM-SDF models inherently allows for runtime prediction of execution paths, a property not readily available or exploited in many other computational models.
  • Exact Workload = Optimal DVS: The ability to determine the exact workload is the key enabler for truly optimal intra-task DVS. By avoiding worst-case assumptions, the system can operate at lower voltages more frequently.
  • Trade-off Management: The method explicitly considers the overhead of the RPI and voltage scaling mechanisms, acknowledging that these processes themselves consume resources and energy, and aims to minimize their impact.

Key Results

The authors evaluated their method on an MPEG4 natural video decoder implemented using an HFSM-SDF model.

  • Energy Consumption Reduction: Compared to a "power down only" approach, their method achieved a 35.5% reduction in energy consumption.
  • Superiority over Existing DVS: When compared to existing DVS methods based on worst-case execution path (WCEP) and average-case execution path (ACEP), their proposed method yielded an additional 20.7% and 12.7% energy reduction, respectively.
  • Voltage Scaling Behavior: Figure 5(b) visually demonstrates that their method applies lower initial voltages compared to WCEP/ACEP methods, reflecting the more accurate workload estimation.

Limitations and Future Work

  • Overhead: The paper acknowledges that the RPI function and voltage scaler introduce runtime, code size, and energy consumption overheads. While code size overhead is reported as negligible (around 4%), runtime and energy consumption overheads are noted as 21.5% and 20.4%, respectively. The authors suggest that minimizing this overhead is crucial for further energy savings.
  • Model Specificity: The method is tailored to the HFSM-SDF model. Its applicability to other computational models or more general programming paradigms might require further investigation.
  • Future Work: The authors implicitly suggest that future work should focus on minimizing the overhead associated with the RPI and voltage scaling mechanisms to achieve even greater energy efficiency.

Summary updated

Abstract

This paper presents a method of intra-task dynamic voltage scaling (DVS) for SoC design with hierarchical FSM and synchronous dataflow model (in short, HFSM-SDF model). To have an optimal intra-task DVS, exact execution paths need to be determined in compile time or runtime. In general programs, since determining exact execution paths in compile time or runtime is not possible, existing methods assume worst/average-case execution paths and take static voltage scaling approaches. In our work, we exploit a property of HFSM-SDF model to calculate exact execution paths in runtime. With the information of exact execution paths, our DVS method can calculate exact remaining workload. The exact workload enables to calculate optimal voltage level which gives optimal energy consumption while satisfying the given timing constraint. Experiments show the effectiveness of the presented method in low-power design of an MPEG4 decoder system.
BibTeX
@inproceedings{lee2003intra,
  author = {Sunghyun Lee and Sungjoo Yoo and Ki‐Young Choi},
  title = {{An intra-task dynamic voltage scaling method for SoC design with hierarchical FSM and synchronous dataflow model}},
  booktitle = {Proceedings of the International Symposium on Low Power Electronics and Design},
  year = {2003},
  doi = {10.1109/lpe.2002.1029555},
}
Figure