ERVIN: Exploring the Optimal Operating Point on Voltage and Precision for Neural Networks Accelerators
The aim of this project is to exploit the intrinsic feature of DNNs, i.e., the capability of DNNs to tolerate bit flips as well as running with acceptable accuracy under low-precision computations. ERVIN provides a framework to analyze these features for DNNs and subsequently, set the best configuration to optimize the energy-accuracy trade-off. This optimal point can be workload- and platform-dependent, both will be studied in ERVIN by covering a large number of DNN benchmarks as well as with various configurations of the underlying hardware platform.
Acknowledgement: ERVIN is granted by OPRECOMP summer of Code 2020.