Neural networks have enabled state-of-the-art approaches to achieve impressive results on many image processing and analysis tasks. However, while gigapixel images are gaining ground in domains like satellite imaging and digital pathology, feeding neural networks directly with these ultra-high-resolution images is still computationally challenging. With a growing number of high-resolution computer vision applications being proposed, the need for an efficient and powerful AI acceleration system targeting gigapixel images rises.

Modern Convolutional Neural Networks (CNNs) are known to be computationally and memory costly owing to the deep structure that is constantly growing. A reconfigurable design is crucial in tackling this difficulty since neural network requirements are always evolving. The suggested architecture is adaptable to the needs of the neural network.

Systolic arrays are critical in parallel computing. They efficiently accomplish tasks like matrix multiplication and signal processing by coordinating a grid of processing components to perform synchronized operations. The structured data flow reduces memory access while increasing processing, resulting in substantial speedups. Systolic arrays are used in a variety of domains, from AI model training to scientific simulations, to improve speed and enable complicated computations that typical sequential approaches struggle with.

The k-Nearest-Neighbours (kNN) algorithm is a popular Machine Learning technique that can be used for a variety of supervised classification tasks. In contrast to other machine learning algorithms which "encode" the knowledge gained from training data to a set of parameters, such as weights and biases, the parameter set of a kNN classifier consists of just labelled training examples. Classification of an unlabelled example takes place by calculating its Euclidean distance (or any other type of distance metric) from all the stored training examples.

Motivation

At SoC Labs, we have need of an accelerator to test our SoC infrastructure and confirmation of our accelerator wrapper design to get size and performance information as well as to try and get ahead and uncover potential problems researchers may experience trying to put their IP into the reference SoC.

Specification

The preliminary design has been broken into two main blocks: