Zero-Aware Low-Precision RNS Scaling Scheme

Scaling is one of the complex operations in Residue Number System (RNS). This operation necessary for RNS-based implementations deep neural networks (DNNs) to prevent overflow. However, state-of-the-art RNS scalers special moduli sets consider 2k modulo as scaling factor, which results a high-precision output with high area and delay. Therefore, low-precision based on multi-moduli factors should be used improve performance. numbers less than scale factor zero output, makes subsequent result faulty. paper first presents formulation hardware architecture four-moduli using new Chinese remainder theorem 2 (New CRT-II) two-moduli factor. Next, scaler circuits are reused achieve minimum overhead. proposed can detect after then transform scaled residues precision when input number not zero.