Booth Fusion: Efficient Bit Fusion Multiplier with Booth Encoding

Abstract

Recently, several attempts have been made to optimize Deep Neural Networks (DNNs) through various hardware acceleration methods. Among them, Bit Fusion, the dynamic bit-level fusion/decomposition hardware architecture, was noted. We introduce a new model structure, Booth Fusion, which makes dynamic bit-level operations more efficient by implementing Bit Fusion with booth encoding. Our design shows improvements in 16.4% for the number of LUT and 14.2% for throughput.