EMAXVR: A programmable accelerator employing near ALU utilization to DSA

The domain specific accelerators (DSAs) have appeared remarkable performances in many real-world applications such as deep learning. However, the benefit on performances is somehow eaten up by the increasing development cost and poor programmability. In this paper, a programmable accelerator is prop...

Full description

Saved in:
Bibliographic Details
Published in2018 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPS) pp. 1 - 3
Main Authors Ichikura, Takahiro, Yamano, Ryusuke, Kikutani, Yuma, Zhang, Renyuan, Nakashima, Yasuhiko
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.04.2018
Subjects
ALU utilization
Convolution
Convolutional neural networks
data reuse
Deep Learning
Hardware
Kernel
Machine learning
Memory management
programmable accelerator
Registers
Online AccessGet full text

Cover

More Information
Summary:The domain specific accelerators (DSAs) have appeared remarkable performances in many real-world applications such as deep learning. However, the benefit on performances is somehow eaten up by the increasing development cost and poor programmability. In this paper, a programmable accelerator is proposed by employing near ALU utilization to DSA, which is an improved version of our previously reported accelerator called EMAXV. As a result, we found our programmable accelerator can compute convolution operations in AlexNet with only 17% lower utilization of ALU and 14% upper rate of data reuse compared with the latest flexible DSA.
ISSN:2473-4683
DOI:10.1109/CoolChips.2018.8373078