Fixed point multi-bit approximate adder based convolutional neural network accelerator for digit classification inference

Approximate computing is a rapidly growing technique to speed up applications with less computational effort while maintaining the accuracy of error-resilient applications such as machine learning and deep learning. Inheritance properties of the machine and deep learning process give freedom for the designer to simplify the circuitry to speed up the computation process at the expense of accuracy of computational results. Fundamental blocks of any computation are adders. In order to optimize it for better performance, 2-bit multi-bit approximate adders (MAPX) are proposed in this work which breaks the lengthy carry chain. In contrast with other approximate larger width adders, instead of using accurate adders for the most significant part, here proposed 2-bit MAPX-1 and MAPX-2 adders are arranged in various ways to compose most and least significant parts. Designed 8-bit and 16-bit adders are evaluated for their performance and error characteristics. Proposed 2-bit MAPX-2 shows better error characteristics whose MED is 0.250 while occupying less area and MAPX-1 consumes less power and delay at the cost of accuracy. Among the extended adders, MAPX 8-bit adder design1 outperforms the best performing APX based 8-bit adder design1. The error performance of it is improved by 14%, 42.1% and 50.4% compared to the existing well-performing APX 8-bit Design1, Design2 and Design3 respectively. Similarly, proposed MAPX 16-bit Design1 exhibits overwhelming performance compared to best performing APX 16-bit Design1, and its error performance is improved by 24.3%, 34.9% and 50.3% compared to APX 16-bit Design1, Design2 and Design3 respectively. In order to evaluate the proposed adder for a real application, extended MAPX 16-bit Design1 is fit in the convolution layer of Low Weights Digit Detector (LWDD) convolutional neural network-based digit classification system. Our modified system accelerates the computation process by 1.25 factors while exhibiting the accuracy of 91% and it best fits error-tolerant real applications. All the adders are synthesized and implemented in the Intel Cyclone IV EP4CE22F17C6N FPGA.