Comparative Analysis Of Inference Performance Of Pretrained Deep Learning Models In Analog Accelerators

Author

Mafizur Rahman, Prairie View A&M University

Date of Award

12-2024

Document Type

Thesis

Degree Name

Master of Science

Degree Discipline

Computer Science

Abstract

Resistive crossbars using non-volatile memory devices have become promising components for implementing Deep Neural Network (DNN) in hardware. However, crossbar-based computations encounter a notable challenge due to device and circuit-level non-idealities. In this study, we explored three recent crossbar simulation tools, namely, AIHWKIT, CrossSim, and MemTorch, and evaluated four pre-trained DNNs for image classification on CIFAR-100 dataset using these tools. All three tools had strong analog simulation functionalities that effectively mimicked the actual hardware environment. To the best of our knowledge, this is the first study using all three simulation tools of analog accelerators, representing three distinctive hardware settings, to evaluate DNN robustness under analog noise and nonlinearities. We first tested the robustness of four DNNs (VGG19, InceptionV3, ResNet50, & MobilenetV2) using different levels of white Gaussian noise as baselines. Then we evaluated their inference performance on the three tools to determine their resilience to analog noise and nonlinearities in hardware environment. Results showed that using CIFAR-100 while all DNNs suffered performance degradation as expected. ResNet50 outperforms outperformed others in two out of three simulators despite real-world hardware

imperfections due to its deep structure. Specifically, the ResNet50 model showed a mere 3% performance drop in CrossSim and, interestingly, a two percent improvement in AIHWKIT. At the same time, InceptionV3 exhibits only a three percent drop from its baseline, while the performance of three other models declined by 12-19% in MemTorch underscoring InceptionV3’s resilience in MemTorch environment. Later, we extended our analysis of the analog DNNs to MNIST and SVHN datasets, finding that ResNet50 and InceptionV3 outperform their digital counterparts on both datasets. This indicated that model design of DNNs played an important role in their resilience to analog noise and nonlinearities, and their inference performance also depends on specific hardware implementation.

Index Terms − Deep Neural Networks, CrossSim, AIHWKIT, MemTorch, Crossbar Simulation

Committee Chair/Advisor

Lin Li

Committee Co-Chair

Lijun Qian

Committee Member

Sherri S. Frizell

Committee Member

Ahmed Abdelmoamen Ahmed

Publisher

Prairie View A&M University

Rights

Date of Digitization

11/22/2024

Contributing Institution

John B Coleman Library

City of Publication

Prairie View

MIME Type

Application/PDF

Recommended Citation

Rahman, M. (2024). Comparative Analysis Of Inference Performance Of Pretrained Deep Learning Models In Analog Accelerators. Retrieved from https://digitalcommons.pvamu.edu/pvamu-theses/1539