Energy Complexity of Convolutional Neural Networks.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Hardware Acceleration

Background:

• Energy efficiency is crucial for deploying convolutional neural networks (CNNs) on low-power mobile devices.
• Existing methods for mapping CNNs to hardware lack machine-independent energy analysis.
• Hardware-specific energy estimations hinder broad exploration of CNN energy optimization.

Purpose of the Study:

• To introduce a simplified, machine-independent theoretical energy complexity model for CNNs.
• To establish a theoretical energy lower bound for CNN computations.
• To derive energy upper bounds for CNN layer evaluation using common data flows.

Main Methods:

• Developed a two-level memory hierarchy model capturing key energy consumption sources.
• Derived analytical expressions for energy lower and upper bounds of CNN layers.
• Validated the model against real hardware implementations using Timeloop/Accelergy on Simba and Eyeriss platforms.

Main Results:

• The proposed theoretical model provides machine-independent energy complexity estimations for CNNs.
• Derived energy lower and upper bounds demonstrate strong asymptotic agreement with empirical data.
• Statistical tests confirm the model's accuracy in predicting energy consumption.

Conclusions:

• The simplified energy complexity model effectively captures CNN energy consumption across diverse hardware.
• This machine-independent approach facilitates energy-optimal CNN mapping for mobile applications.
• The validated model aids in designing energy-efficient hardware accelerators for deep learning.