T., Glint, TomGlint, TomT.C.K., Jha, Chandan KumarJha, Chandan KumarC.K.M., Awasthi, ManuAwasthi, ManuM.J., Mekie, JoyceeMekie, JoyceeJ.2025-09-012025-09-019.80E+1210.1109/ISPASS57527.2023.000492-s2.0-85164540093https://www.scopus.com/inward/record.uri?eid=2-s2.0-85164540093&doi=10.1109%2FISPASS57527.2023.00049&partnerID=40&md5=d2f0cf0f2e2082f914b98ec371344be2https://d8.irins.org/handle/IITG2025/29396Various DNN accelerators based on Conventional compute Hardware Accelerator (CHA), Near-Data-Processing (NDP) and Processing-in-Memory (PIM) paradigms have been proposed to meet the challenges of inferencing Deep Neural Networks (DNNs). To the best of our knowledge, this work aims to perform the first quantitative as well as qualitative comparison among the state-of-the-art accelerators from each digital DNN accelerator paradigm. Our study provides insights into selecting the best architecture for a given DNN workload. We have used workloads of the MLPerf Inference benchmark. We observe that for Fully Connected Layer (FCL) DNNs, PIM-based accelerator is 21� and 3� faster than CHA and NDP-based accelerator respectively. However, NDP is 9� and 2.5� more energy efficient than CHA and PIM for FCL. For Convolutional Neural Network (CNN) workloads, CHA is 10% and 5� faster than NDP and PIM-based accelerator respectively. Further, CHA is 1.5� and 6� more energy efficient than NDP and PIM-based accelerators respectively. � 2023 Elsevier B.V., All rights reserved.EnglishConvolutional neural networksData handlingEnergy efficiencyConvolutional neural networkData processorsDeep neural network acceleratorEnergy efficientHardware acceleratorsMemory paradigmNear data processorNeural-network processingProcessing-in-memoryState of the artDeep neural networksConference paper20233