# Design Of High Accuracy Fixed Width Modified Booth Multiplier For Mac

The proposed approach can be integrated into existing CNN accelerators, e.g., [7,9,20,21,22]. However, to adopt the proposed approach, their processing element (PE) designs should be modified to deal with nine pixels at the same time. Note that, in the original PE designs of these CNN accelerators [7,9,20,21,22], their convolution operations are performed by multipliers and adders. If the proposed approach is adopted, their convolution operations are accelerated.

## Design Of High Accuracy Fixed Width Modified Booth Multiplier For Mac

Our PPM is a modification to the typical PPM of radix-4 booth encoding [30]. Without loss of generality, here we use 8-bit multiplication as an example for illustration. Figure 6 gives the typical PPM of radix-4 booth encoding for signed 8-bit multiplication. As displayed in Figure 6, the typical PPM (of radix-4 booth encoding) includes signed extension terms, normal partial product (PP) terms, least significant bit (LSB) term, and negative carry-in (Neg_cin) terms. It should be mentioned that the Boolean expressions of these terms, including signed extension terms, normal PP terms, LSB terms, and Neg_cin terms, have been elaborated in [32]. Moreover, according to [30], we can modify the Boolean expressions of both LSB terms and Neg_cin terms, and then, we can move each Neg_cin term to one higher bit position, as shown in Figure 7. Compared with previous PPMs [30,32], this PPM (i.e., Figure 7) does have any accuracy loss. Note that our following discussions and optimizations are based on this PPM (i.e., Figure 7).