Bilkent University
Department of Computer Engineering

---

CS 590/690 SEMINAR

 

Efficiency Aware Reconfigurable CNN Accelerator

 

Alperen Kalay
Master Student
(Supervisor: Prof.Dr.Özcan Öztürk)
Computer Engineering Department
Bilkent University

Abstract: Accelerators are specialized hardwares for target applications. Convolutional Neural Network (CNN) accelerators have become increasingly vital in the realm of deep learning. They aim to optimize dataflows for convolution operation, often employing systolic arrays with mesh topology architectures. The efficiency of these accelerators is inherently tied to the selection of dataflow techniques based on loop ordering, each presenting its unique advantages and disadvantages. In this work, efficiency aware reconfigurable CNN accelerator is designed on a systolic array. Reconfigurability helps to mitigate idle cores by dynamically adapting suitable dataflow techniques. However, highly reconfigurable architectures require dynamicality at extreme levels which weaken performance. The primary objective of this research is to maximize efficiency while minimizing reconfigurability. It is achieved through the combination of two dataflows with other unique design criterias. The paper delves into details of the proposed architecture, discussing its position among different trade-off planes to provide a comprehensive understanding of its design choices.

 

DATE: March 18, Monday @ 14:10 Place: EA 502