返回
MCPS: a mapping method; MAERI accelerator base on Cartesian Product based Convolution; DNN layers with sparse input feature map
DOI:10.1007/s10586-021-03527-6.png)
摘要
En 中文
Today, the high accuracy of deep learning has led to use in various domains such as image and voice classification. However, vast computations of deep neural networks (DNNs) have caused the inefficiency of traditional processors, resulting in the emergence of hardware accelerators. DNN accelerators have increased per;
mance by exploiting opportunities such as data reuse and sparsity. In these accelerators, the dataflow is an essential factor, so that some of them have a reconfigurable architecture to support different mappings and dataflows. However, the accelerators explicitly designed;
exploiting the data sparsity are usually non-reconfigurable and have fixed dataflow. This paper presents a new dataflow called Channel Dimension Stationary (CDS);
the MAERI (a Reconfigurable Neural Network Accelerator). It can be used;
convolutional layers with sparse input feature maps (ifmaps). In the proposed dataflow, computations are based on the Cartesian product method. However, multiplications leading to useless results are avoided. To analyze the mapping based on CDS dataflow, we upgraded the mRNA tool (mapper;
Reconfigurable Neural Accelerators), which includes an energy and per;
mance analyzer of the mapping strategy in MAERI. By evaluation, we found that in the sparse ifmaps of 50%, 70%, and 90%, the proposed mapping on average can increase energy efficiency by 3x, 6x, and 13x respectively, without noticeable reduction of utilization.
Keyword:
Cartesian Product based Convolution (CPC)
Channel Dimension Array (CDA)
Convolutional layer
Input feature map(ifmap)
Mapping
Sparsity
期刊
C
IF:
4.1
论文数:
5.1K
被引数:
7.5K
机构
引用论文
Structural study of lanthanides(III) in aqueous nitrate and chloride solutions by EXAFS通过EXAFS对硝酸盐和氯化物水溶液中镧系元素 (III) 的结构研究

