Energy Profiling of USB DNN Accelerators
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
DescriptionWe measure power and energy consumption of the NCS2 and the Edge TPU for Neural Networks. We investigate how accuracy decreases with lower sampling frequency and how well number of operations and latency can serve as estimates for energy consumption for individual layers. Our main conclusions are that a sampling frequency of 200 kHz is required for a minimal accuracy of 5 %; latency is a much better estimator than the number of operations with an expected accuracy of 10% for overall energy consumption; specific hardware settings may have significant influence and must be considered in a reliable energy estimation.