Mantra I: A Systolic Array for Neural Computation

Abstract

In the last decade, the need has arisen for medium-cost dedicated computers for artificial neural network (ANN) models. Several machines have been proposed. However, only very seldom can systems be considered as massively parallel and, hence, exploit the huge intrinsic parallelism of ANN models. The MANTRA I machine addresses this issue by targeting synapse-level parallelism on a bidimensional systolic array, based on a custom VLSI circuit called GENES IV. A prototype SIMD computer with 400 processing elements (PEs), expandable to 1600 PEs, has been designed, built, tested, and analyzed.This chapter focuses mainly on the architecture of the machine. The performance of the machine is analyzed using the Kohonen ANN model. Finally, a critical analysis suggests how the same general approach could be kept for future systems, requiring important steps toward generality and better ease-of-use. IntroductionResearch on artificial neural networks (ANNs) and computer science have coexisted for almost six decades. In the 80's, the ANN domain 71 Neural Networks and Systolic Array Design Downloaded from www.worldscientific.com by MONASH UNIVERSITY on 04/23/17. For personal use only.

Self-organization of the batch Kohonen network under quantization effects

“…The analysis performed in this paper includes as a special case (for a unit epoch length), the original algorithm studied in [35]. The advantage of the batch variant is that it decorrelates the winner selection and weight update equations inside an epoch making batch processing possible which can then be exploited by massively parallel -at the synaptic level -machines such as systolic array architectures [15,39].…”

“…the T winners during the kth epoch depend not only on the inputs ξ(kT + t ) but also on the actual weight values μ i (kT + t ) and, therefore, cannot be precomputed. To be able to utilize special parallel hardware, such as the mantra machine [15,39], appropriate variants of Kohonen's algorithm that permit decoupling of Equations (3) and (5) have been proposed in [16]. The variant studied in this work is the batch version of Kohonen's algorithm [16,36] in which weight changes are accumulated throughout each epoch and the weights are modified only after the presentation of the last input ξ(kT + T − 1).…”