Using Libraries of Approximate Circuits in Design of Hardware Accelerators of Deep Neural Networks

Abstract

Approximate circuits have been developed to provide good tradeoffs between power consumption and quality of service in error resilient applications such as hardware accelerators of deep neural networks (DNN). In order to accelerate the approximate circuit design process and to support a fair benchmarking of circuit approximation methods, libraries of approximate circuits have been introduced. For example, EvoApprox8b contains hundreds of 8-bit approximate adders and multipliers. By means of genetic programming we generated an extended version of the library in which thousands of 8- to 128-bit approximate arithmetic circuits are included. These circuits form Pareto fronts with respect to several error metrics, power consumption and other circuit parameters. In our case study we show how a large set of approximate multipliers can be used to perform a resilience analysis of a hardware accelerator of ResNet DNN and to select the most suitable approximate multiplier for a given application. Results are reported for various instances of the ResNet DNN trained on CIFAR-10 benchmark problem.