EUVIMEDEUVIMED entdecken
Crossref

Energy Optimization for FPGA-Based CNN And SVM Accelerator: A Comparative Design Framework

Manish Kumar, Manoj Pandey, Bandi Nageshwar Rao

IOSR Journal of VLSI and Signal processing · 2026-05 · 16 (5) · 22-33

DOI: 10.9790/4200-16052233 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IOSR Journal of VLSI and Signal processing, 2026-05. Verfasst von Manish Kumar und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The research explores various hardware acceleration approaches, beginning with Graphics Processing Units (GPUs) that leverage their massively parallel architecture to accelerate matrix operations fundamental to neural networks. We examine the architectural advantages of GPUs, including their thousands of cores optimized for floating-point operations, high memory bandwidth, and specialized tensor processing units. The analysis extends to Application-Specific Integrated Circuits (ASICs) such as Google's Tensor Processing Units (TPUs), which represent a paradigm shift toward domain-specific architectures optimized exclusively for machine learning workloads. These specialized processors demonstrate significant improvements in performance per watt and throughput compared to general-purpose alternatives. The paper analyses key performance metrics including computational throughput measured in operations per second, energy efficiency expressed as operations per joule, memory bandwidth utilization, and latency characteristics across different acceleration platforms. Real-world benchmarking studies demonstrate that specialized hardware can achieve 10-100x performance improvements over traditional CPUs for common deep learning tasks, with energy efficiency gains often exceeding 1000x for inference workloads. Field-Programmable Gate Arrays (FPGAs) are investigated as a flexible middle ground, offering reconfigurable hardware that can be optimized for specific neural network architectures while maintaining adaptability for evolving ML algorithms. The study examines how FPGAs enable custom data path designs, variable precision arithmetic, and pipeline optimization that can significantly reduce inference latency and energy consumption. Additionally, we explore emerging technologies including neuromorphic computing architectures that mimic biological neural networks, in-memory computing solutions that eliminate the von Neumann bottleneck, and quantum computing approaches for specific ML applications.

Zitieren

APA

Kumar, M., Pandey, M., & Rao, B. N. (2026). Energy Optimization for FPGA-Based CNN And SVM Accelerator: A Comparative Design Framework. IOSR Journal of VLSI and Signal processing, 16(5), 22-33. https://doi.org/10.9790/4200-16052233

Vancouver

Kumar M, Pandey M, Rao BN. Energy Optimization for FPGA-Based CNN And SVM Accelerator: A Comparative Design Framework. IOSR Journal of VLSI and Signal processing. 2026;16(5):22-33. doi:10.9790/4200-16052233

BibTeX
@article{Kumar2026,
  author = {Kumar, Manish and Pandey, Manoj and Rao, Bandi Nageshwar},
  title = {Energy Optimization for FPGA-Based CNN And SVM Accelerator: A Comparative Design Framework},
  journal = {IOSR Journal of VLSI and Signal processing},
  year = {2026},
  volume = {16},
  number = {5},
  pages = {22--33},
  doi = {10.9790/4200-16052233}
}
Kostenlos starten →