DICHA - DNN based Intelligent classification for Workload Characterization on Heterogeneous Architecture

Abstract

Nowadays Heterogeneous system on-chip (HSoC) become highly essential.  IoT, Industry 4.0, intelligent vehicles, embedded devices, and cyber-physical framework applications are broadly utilizing such equipment models for workload processing. These continuous applications include a miscellaneous set of workloads with various attributes which highly influences the computational cycles. Moreover, asset organization become a basic issue in HSoC. In this paper DNN based classifier is proposed for HSoC stages to predict ideal computational asset for every responsibility at runtime. Deep Neural Networks (DNN), with deep layers and extremely high element of boundaries, have exhibited get through learning capacity in Machine learning region. Nowadays DNN with Big Data input are driving another heading in enormous scope object acknowledgment. The proposed classifier analysed the execution of a few HSoC stages to comprehend the functioning guideline of ongoing responsibilities at runtime. The noticed attributes are outlined as continuous data set and the equivalent is used to train and test the DNN classifier. The proposed classifier is assessed on raspberry-pi HSoC and re-enacted on the python with ML library. Precision, throughput, affectability, selectivity measurements are distinguished to break down the exhibition of the proposed calculations. The proposed DICHA framework accomplished the precision up to 96% contrasted and outrageous ML predictor for and furthermore saved the execution energy up to 30% for real-time embedded benchmark workloads like MiBench, IoMT.