Cellular Processor Arrays

Members: Jorg Conradt, Ralph Etienne-Cummings, Steve Kalik, Tobi Delbruck

Cellular Processor Arrays integrate thousands of very simple processing ‘cores’ in a single chip. These can be effectively used to perform computations at high speed and with lower power consumption as compared with conventional processor systems. The devices commonly use a SIMD (Single Instruction Multiple Data) processing model, where individual processors all perform the same instructions, while communication is localised to nearest neighbourhood. This programming model is particularly useful when models consist of regular arrays of data, such as low-level image processing, neural models based on topographic maps and sheets of neural layers, dynamic neural fields models, and similar.

We will run tutorials on SIMD programming, developing algorithms on the SCAMP-3 system - a ‘vision chip’ with 128x128 processor array (we brought a few boxes), and on using the APRON array processing software for general array-based computations, in particular for constructing neural models. The workgroup may also discuss GPU-based computation (CUDA etc.), which follows a similar programming model. Further tutorials will be determined depending on the interest, and relation to the projects, e.g. on circuit design of the ‘analogue microprocessor’, on interfacing your own hardware/software to the APRON framework (writing your own dll’s) etc.

A few projects last year made use of the array processing, and we hope this year the devices/techniques explored in this group could be used in a number of projects groups, e.g. wiki:2010/sf10, wiki:2010/att10, wiki:2010/rob10

We brought a few SCAMP boxes with us, and everybody is welcomed to download and use the APRON software, so if you think you may be interested in using these in your project, have some new ideas you would like to explore, or just want to learn more how to use these technologies and have some hands-on experience of programming a 16,384-processor chip, then join our tutorials.

Some project ideas: - Visual pre-processing for multi-modal data fusion and learning of internal representations - Action-selection with bottom-up saliency map formation and inhibition-of-return on a single chip - On-sensor vision processing for autonomous robot navigation - Cellular processor arrays in the context of event-based processing: using the vision chip to produce complex spatio-temporal events. - Implementation of a switched-current analogue processor on the FPAA device - Implementing arrays of spiking neurons on the SCAMP chip - etc.