Bit-Beading: Stringing bit-level MAC results for Accelerating Neural Networks

On account of the rising demands of AI applications and the consequent need for improvement, researchers are designing better and faster algorithms and architectures. Convolutional Neural Networks (CNN) are neural networks that have become ubiquitous and find applications in the domain of computer vision. Inference in CNN involves convolution operation, which mainly consists of a massive number of matrix multiplications. Optimising these multiplications will enable faster execution of the inference tasks. Fixed precision during inference takes the same time to compute for both higher and lower precision. It is noted in the literature that lowering the precision to some extent does not affect the inference accuracy. In this paper, we propose a reconfigurable multiplier that can handle the precision of different magnitudes. We design Bit-Bead, a basic unit based on Booth's algorithm, where several bit-beads are composed (i.e., stringed) to form a multiplier of the required precision. The reconfigurable multipliers need low latency due to lower precision and also enable performing multiple low-precision computations. Our proposal shows considerable performance improvement compared to the baseline and existing designs.