1. Introduction
This paper is published in VLSI in 2017.
Prior works on DCNN acceleration usually use a fixed computation pattern for diverse DCNN models, which leads to imbalance between power efficiency and performance. Thus, the researchers design a DCNN acceleration architecture, called deep neural architecture (DNA), that aims to sovle this problem.
2. Innovation points
The following is the general architecture of DNA.
1. DNA is able to reconfigure its data paths to support a hybrid data reuse pattern (input reuse, output reuse and weight reuse) for different layer sizes.
2. DNA is able to reconfigure its computing resources by adopting a parallel output oriented mapping method.
3. Propose a layer-based scheduling framework.
3. Summary
The DNA's layout and characteristics are shown below.