Accelerating Solar Panel Snow Detection: CPU vs. GPU Training Performance of VGG-19 and ResNet-50
Year: 2026
Type: article
Abstract: Solar panel snow detection is critical for maintaining photovoltaic system efficiency in cold climates. While deep learning models such as VGG-19 and ResNet-50 achieve near-perfect classification accuracy for this task, their training demands substantial computational resources. This study presents a systematic performance analysis of CNN training on the TRUBA national high-performance computing infrastructure, comparing multi-threaded CPU parallelization (1–112 threads on Intel Xeon Platinum 8480+) with GPU acceleration (NVIDIA V100). Experiments reveal that the model architecture fundamentally determines CPU scaling behavior: VGG-19, dominated by large fully connected layers, achieves 28× speedup at 56 threads (serial fraction f ? 1.8%), whereas ResNet-50, composed of many small residual operations, scales to only 1.56× ( f ? 63.4%). GPU training provides 163–1067× acceleration over single-threaded CPU execution for both architectures. Classification quality (accuracy and F1-score ? 0.995) remains invariant across all hardware configurations, confirming that training infrastructure selection is purely a performance optimization decision.
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7214531500