SCDVP: A Simplicial CNN Digital Visual Processor

Abstract

In this work we present a programmable and reconfigurable single instruction multiple data (SIMD) visual processor based on the S-CNN architecture, namely, the Simplicial CNN Digital Visual Processor (SCDVP), oriented to high-performance lowlevel image processing. The cells in the array have a selectable neighborhood configuration and several registers, which provide the chip with extended spatial and temporal processing capabilities, in particular optical flow. A prototype 64 64 cell chip with two program memories and a column adder was fabricated in a 90 nm technology, which running at 133 MHz delivers 105.5 GOPS. The calculation at the cell level is performed with time coded signals and the program memory is located outside the array. This produces a very efficient realization in terms of area: 53.8 GOPS per mm , which outperforms all results reported so far. We show that even after normalization, to account for technology scaling, the proposed architecture is the most efficient among all reported digital processors. Computation performance to power ratio also exceeds all previous results with 817.8 GOPS/W. Experimental results of the working chip are reported.

Simplicial computation: A methodology to compute vector–vector multiplications with reduced complexity

“…This requires comparisons of q bits and then N subtractions of q bits. The second option that turns out to be much more efficient from the energy/area viewpoint (extensively used in compact realizations like other studies 30,31 20 ) consists of encoding inputs in time by comparing them with an increasing ramp signal, as illustrated in Figure 11A. The ramp always has Q clock cycles, regardless of the number of inputs N .…”

“…In order to show this, we compare the corresponding digital architectures and evaluate both at a fundamental level, in terms of the number of standard CMOS gates and operation cycles. In addition, we propose a modified comparator for the simplicial algorithm that minimizes activity and produces a significant reduction in energy with respect to previous realizations 20,21 …”

On the Impact of Energy-Accuracy Tradeoff in a Digital Cellular Neural Network for Image Processing

“…Analog [12], mixed-mode [13], and digital [14], [15] CNN designs have been presented. Analog and mixed-mode (analog arithmetic but digital memory) have better performance but difficult to program and less scalable.…”

QoE-driven big data management in pervasive edge computing environment

“…Compared with traditional methods, machine learning [10,11] techniques have some unique advantages in the extraction of big data and many studies have applied deep learning [12,13] techniques in the pervasive edge computing environment. A typical example of a straightforward solution to achieving state-of-the-art accuracy in high-dimensional big data analysis is the use of a Convolutional Neural Network (CNN) technique [14,15] , such as image/video processing, speech recognition, or natural language processing. Another option is to use Deep CNN (DCNN) [16] to perform high-dimensional big data analysis, which yields higher accuracy than CNN.…”

“…In general, Refs. [14][15][16][17] have presented solutions for gradually improving accuracy, and the authors in Ref. [18] were able to increase the training speed using CNN.…”