Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CarbonCP: Carbon-Aware DNN Partitioning with Conformal Prediction for Sustainable Edge Intelligence

Hongyu Ke, Wanxin Jin, Haoxin Wang

TL;DR

The paper tackles the problem of high operational carbon emissions in end-to-end edge intelligence by proposing CarbonCP, a context-adaptive, carbon-aware, and uncertainty-aware DNN partitioning framework built on conformal prediction. It introduces an implicit predictor $Q_{\theta}$ to estimate partition performance, a CP-based score function, and a weighted CP scheme to handle covariate shift between calibration and test contexts. The approach yields prediction intervals for the optimal partition with statistically guaranteed coverage and enables three decision modes (R, M, A), with CarbonCP-A delivering the best carbon-footprint reductions in dynamic settings. Experimental results on a ResNet-18 deployment demonstrate substantial carbon reductions (up to $58.8\%$) while maintaining competitive latency and energy metrics, and show that carbon intensity and edge resource usage are critical levers for improvement.

Abstract

This paper presents a solution to address carbon emission mitigation for end-to-end edge computing systems, including the computing at battery-powered edge devices and servers, as well as the communications between them. We design and implement, CarbonCP, a context-adaptive, carbon-aware, and uncertainty-aware AI inference framework built upon conformal prediction theory, which balances operational carbon emissions, end-to-end latency, and battery consumption of edge devices through DNN partitioning under varying system processing contexts and carbon intensity. Our experimental results demonstrate that CarbonCP is effective in substantially reducing operational carbon emissions, up to 58.8%, while maintaining key user-centric performance metrics with only 9.9% error rate.

CarbonCP: Carbon-Aware DNN Partitioning with Conformal Prediction for Sustainable Edge Intelligence

TL;DR

The paper tackles the problem of high operational carbon emissions in end-to-end edge intelligence by proposing CarbonCP, a context-adaptive, carbon-aware, and uncertainty-aware DNN partitioning framework built on conformal prediction. It introduces an implicit predictor to estimate partition performance, a CP-based score function, and a weighted CP scheme to handle covariate shift between calibration and test contexts. The approach yields prediction intervals for the optimal partition with statistically guaranteed coverage and enables three decision modes (R, M, A), with CarbonCP-A delivering the best carbon-footprint reductions in dynamic settings. Experimental results on a ResNet-18 deployment demonstrate substantial carbon reductions (up to ) while maintaining competitive latency and energy metrics, and show that carbon intensity and edge resource usage are critical levers for improvement.

Abstract

This paper presents a solution to address carbon emission mitigation for end-to-end edge computing systems, including the computing at battery-powered edge devices and servers, as well as the communications between them. We design and implement, CarbonCP, a context-adaptive, carbon-aware, and uncertainty-aware AI inference framework built upon conformal prediction theory, which balances operational carbon emissions, end-to-end latency, and battery consumption of edge devices through DNN partitioning under varying system processing contexts and carbon intensity. Our experimental results demonstrate that CarbonCP is effective in substantially reducing operational carbon emissions, up to 58.8%, while maintaining key user-centric performance metrics with only 9.9% error rate.
Paper Structure (20 sections, 22 equations, 10 figures, 1 algorithm)

This paper contains 20 sections, 22 equations, 10 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminary Experiments and Key Insights
  3. The Impact of Dynamic Processing Context on DNN Partitioning
  4. The Impact of Carbon Intensity on DNN Partitioning
  5. Lack of Uncertainty Assessment in DNN Partitioning
  6. Conformal Prediction
  7. Problem Statement
  8. CarbonCP Design
  9. Implicit Predictor Model
  10. Score Function
  11. Confidence Interval
  12. Assess Prediction Interval Performance
  13. Performance Evaluation
  14. Dataset Construction
  15. Performance Comparison
  16. ...and 5 more sections

Figures (10)

  • Figure 1: Trade-offs among the operational carbon emissions of the end-to-end edge computing system, end-to-end latency, and battery consumption. Lower is better.
  • Figure 2: Impact of dynamic processing contexts on DNN partitioning solutions, achieving minimal overall operational carbon emissions vs. minimal end-to-end latency in edge computing systems.
  • Figure 3: Carbon intensity (CI) varies both spatially and temporally in the United States Emaps.
  • Figure 4: Impact of carbon intensity on DNN partitioning.
  • Figure 5: Overview of the proposed CarbonCP framework.
  • ...and 5 more figures