**Author details**

we can consider that the power consumption of DRAM access is the main part of power consumption of memory access. Storing and reusing data with on-chip buffer can reduce the DRAM access, so that it can reduce the power consumption of the computing process of CNN

As **Figure 15** shown, the off-chip memory access of point B is 4.5 times more than point A, while the on-chip buffer size of point B is only 2 KB approximately more the point A. It means we can increase a small on-chip buffer (2 KB) to obtain the 4.5 times reduction of the power

Although it is theoretical estimate of power consumption, it shows that we can make a balance between on-chip memory size and power consumption. Our design space evaluation method can help us to choose the optimal design space of CNN accelerator which can reduce the power consumption by increasing small on-chip buffer size. Facing the tendency that miniaturization and low power consumption of IOT, our evaluation method is an effective design

Convolution neural network (CNN) has been developing rapidly and used widely for many computer science domains in the past few years, such as image recognition, speech recognition, game play, etc. In the image recognition filed, the recognition accuracy of ResNet exceeded human accuracy in 2015. The outstanding performance makes CNN more and more popular in the artificial intelligence applications. Many researches increase the depth of CNN model to improve the accuracy; in the meanwhile, it brings great pressure to the hardware. Therefore, many specific CNN accelerators are designed and used for CNN computing, including FPGA designs, GPU designs, and ASIC designs, which aim to improve the comput-

In this chapter, we reviewed the history of CNN and introduce the basic and principle of CNN. Following we presented a real-life CNN model, namely VGG-16. We illustrated several CNN accelerators and then we introduced and analyzed two optimization methods of CNN accelerators, including reducing data precision and data-reusing. Based on the analysis, we enumerated all legal design possibilities of CNN accelerator and the optimal design space for CNN accelerator can be obtained. By depicting the design space groups on a graph, we obtained the appropriate design space of CNN accelerator according to our design

This work was supported by the Project Science and Technology of Guangdong Province of

ing performance and reduce the energy consumption of CNN computing.

China (2015B090912001, 2016B010123005, and 2017B090909005).

accelerator.

164 Green Electronics

**5. Summary**

requirements.

**Acknowledgements**

consumption of memory access of CONV1.

strategy and match the concept of green electronics.

Wenquan Du, Zixin Wang and Dihu Chen\*

\*Address all correspondence to: stscdh@mail.sysu.edu.cn

Sun Yat-Sen University, Guangzhou, China
