Strategic Improvements in CNN Accelerators: Optimizing PE Utilization for MobileNetV2
The continuous evolution of deep neural networks (DNNs) demands more efficient hardware implementations for real-time applications on mobile and edge devices. This paper extends previous research by enhancing the optimization of Processing Elements (PEs) and introducing a novel data flow for the calculation of Depthwise Separable Convolution (DSC) in MobileNetV2. Our upgraded processing element hardware incorporates advanced architectural modifications that substantially improve throughput and computational efficiency.
Authors
- Sang-Seol Lee (ORCID: https://orcid.org/0009-0001-5158-438X)
- M. y. Lee (ORCID: https://orcid.org/0000-0002-1347-1050)
- Eunchong Lee (ORCID: https://orcid.org/0000-0001-7873-2920)
- Sung‐Joon Jang (ORCID: https://orcid.org/0000-0002-6992-408X)
- Kyungho Kim (ORCID: https://orcid.org/0000-0002-9717-6867)
Institutions
- Korea Electronics Technology Institute (KR)
Publication Details
- Published
- 2024-08-19
- DOI
- https://doi.org/10.1109/isocc62682.2024.10762664
- Primary Topic
- Robotics and Automated Systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00