Tiled architecture of a CNN-mostly IP system

Lambert Spaanenburg; Suleyman Malki

doi:10.1117/12.821904

28 May 2009 Tiled architecture of a CNN-mostly IP system

Lambert Spaanenburg, Suleyman Malki

Proceedings Volume 7363, VLSI Circuits and Systems IV; 736313 (2009) https://doi.org/10.1117/12.821904
Event: SPIE Europe Microtechnologies for the New Millennium, 2009, Dresden, Germany

Abstract

Multi-core architectures have been popularized with the advent of the IBM CELL. On a finer grain the problems in scheduling multi-cores have already existed in the tiled architectures, such as the EPIC and Da Vinci. It is not easy to evaluate the performance of a schedule on such architecture as historical data are not available. One solution is to compile algorithms for which an optimal schedule is known by analysis. A typical example is an algorithm that is already defined in terms of many collaborating simple nodes, such as a Cellular Neural Network (CNN). A simple node with a local register stack together with a 'rotating wheel' internal communication mechanism has been proposed. Though the basic CNN allows for a tiled implementation of a tiled algorithm on a tiled structure, a practical CNN system will have to disturb this regularity by the additional need for arithmetical and logical operations. Arithmetic operations are needed for instance to accommodate for low-level image processing, while logical operations are needed to fork and merge different data streams without use of the external memory. It is found that the 'rotating wheel' internal communication mechanism still handles such mechanisms without the need for global control. Overall the CNN system provides for a practical network size as implemented on a FPGA, can be easily used as embedded IP and provides a clear benchmark for a multi-core compiler.

Citation Download Citation

Lambert Spaanenburg and Suleyman Malki "Tiled architecture of a CNN-mostly IP system", Proc. SPIE 7363, VLSI Circuits and Systems IV, 736313 (28 May 2009); https://doi.org/10.1117/12.821904

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available