Achieve Energy Efficiency with Axis System for Edge Devices

Achieve Energy Efficiency with Axis System for Edge Devices

Table of Contents

1. Introduction
2. Deep Learning at the Edge
3. Energy Efficiency in DNA Inference
4. Approximate Computing in Edge Devices
5. Approximation in Different Subsystems
6. The Axis System
   • Introducing the Axis System
   • Designing Cam Edge and Cam Cloud
   • Prototype Implementation and Evaluation
7. Comparison of Axis with Individual Approximations
8. Landscape of Approximations for DNNS
9. Current Research at Intel AI
10. Conclusion

Introduction

In recent years, there has been a growing interest in leveraging artificial intelligence (AI) for various applications, ranging from smart cities and autonomous driving to virtual reality and Speech Recognition. This has led to the development of deep learning techniques that allow AI models to be deployed on edge or end user devices. However, these low-power devices with limited compute, memory, and battery life Present challenges in terms of computational complexity and energy consumption.

Deep Learning at the Edge

Deep learning at the edge refers to the use of AI models on edge devices such as smartphones, smart cameras, and smart home devices. These devices enable applications like speech recognition, face recognition, object detection, and Image Segmentation. The goal of deep learning at the edge is to run these AI models efficiently on low-power devices powered by batteries.

Energy Efficiency in DNA Inference

Energy efficiency is a crucial requirement for edge devices. Deep neural networks (DNNs) have demonstrated high accuracy levels for various visual analytics applications. However, they come with challenges such as high computational complexity and energy consumption. To achieve sustainable and green Edge AI, extreme energy efficiency in DNA inference algorithms is needed.

Approximate Computing in Edge Devices

Approximate computing is a technique that leverages application resilience to achieve energy benefits. This technique has been widely used and studied in the field. In the context of edge devices, approximate computing can play a vital role in achieving energy efficiency in DNA inference systems.

Approximation in Different Subsystems

To fully leverage DNN error resiliency and extract maximum energy efficiency, it is essential to explore approximation not only in the compute part but also in other subsystems of edge devices. This includes approximation techniques for the sensor subsystem, memory subsystem, compute part, and communication subsystem. By approximating the entire system, rather than just the compute part, optimal energy efficiency can be achieved.

The Axis System

The Axis system is a proposed approximate inference system for energy-efficient inference at the edge. It introduces approximation in different subsystems and is designed to optimize energy efficiency. The system consists of two variants: Cam Edge and Cam Cloud.

Introducing the Axis System

The Axis system is the first and only work that proposes an approximate inference system for energy-efficient inference at the edge. It incorporates approximation techniques in different subsystems to achieve energy savings.

Designing Cam Edge and Cam Cloud

Cam Edge and Cam Cloud are two variants of the Axis system. They are designed to optimize energy efficiency in edge devices for different use cases. The system has been prototyped and evaluated with various applications, including classification, object detection, and segmentation.

Prototype Implementation and Evaluation

A prototype of the Axis system has been implemented on an Intel FPGA board. The evaluation includes 14 different neural networks and a comparison of the performance of Axis with individual subsystem approximations. The results show up to 4X energy savings with the Axis system.

Comparison of Axis with Individual Approximations

A comparison is made between the Axis system and individual subsystem approximations for energy savings. The results demonstrate that the Axis system achieves better energy efficiency compared to approximating individual subsystems.

Landscape of Approximations for DNNS

The landscape of approximations for DNNs includes hardware, algorithmic, software, and hardware-software co-approximations. Techniques such as quantization, voltage scaling, and refresh reduction are used to optimize energy efficiency in DNNs.

Current Research at Intel AI

In the Advanced Architecture Research group at Intel AI, current research focuses on enabling energy-efficient AI using approximations. The impact of different approximations, such as quantization, on DNN performance is being explored. The research aims to increase the speed and energy efficiency of AI hardware and software design.

Conclusion

The concept of the approximate inference system, such as the Axis system, empowers developers to build energy-efficient and smart systems. By optimizing all subsystems of edge devices, the best energy efficiency can be achieved. This enables the widespread availability of AI to make a significant impact in the world.

References:

1. Reference 1
2. Reference 2
3. Reference 3
4. Reference 4
5. Reference 5