Systolic VLSI and FPGA Realization of Artificial Neural Networks

Abstract

Systolic architectures are established as a widely popular class of VLSI structures for repetitive and computation-intensive applications due to the simplicity of their processing elements (PEs), modularity of design, regular and nearest neighbor interconnections between the PEs, high-level of pipelinability, small chip-area and low-power consumption. In systolic arrays, the desired data is pumped rhythmically in a regular interval across the PEs to yield high throughput by fully pipelined processing. The I/O bottleneck is significantly reduced by the systolic array architectures by feeding the data at the chip-boundary, and pipelining it across the structure. The extensive reuse of data within the array allows for executing large volume of computation with only a modest increase of bandwidth. Since the FPGA devices consist of regularly placed inter-connected logic blocks, they closely resemble with the layout of systolic processors. The systolic computation within the PEs therefore could easily be mapped to the configurable logic blocks in FPGA device. Interestingly also, the artificial neural network (ANN) algorithms are quite suitable for systolic implementation due to their repetitive multiply-accumulate behaviour. Several variations of one-dimensional and two-dimensional systolic arrays are, therefore, reported in the literature for the implementation of different types of neural networks. Special purpose systolic designs for various ANN-based applications relating to pattern recognition and classification, adaptive filtering and channel equalization, vector quantization, image compression and general signal/image processing applications have been reported in the last two decades. We have devoted this chapter on the systolic architectures for the implementation of ANN algorithms in custom VLSI and FPGA platforms. The key techniques used for the design of basic systolic building blocks of ANN algorithms are discussed in detail. Moreover, the mapping of fully-connected unconstrained ANN, as well as, multilayer ANN algorithm into fully-pipelined systolic architecture is described with generalized dependence graph formulation. A brief overview of systolic architectures for advance ANN algorithms for different applications are presented at the end.