CNN Accuracy–Efficiency Trade-offs Across Computing Platforms: A Literature-Motivated Controlled Deployment Experiment

Abstract

The most popular architecture for land-cover classification and remote sensing remains image classification, and EuroSAT remains the benchmark dataset. Six studies from the literature (2023–2026) report either classification accuracy or deployment efficiency for CNNs, but none report both dimensions for the same model across multiple platforms in an image-classification setting. This paper closes that gap with a controlled deployment experiment. A single ResNet-18 was trained once on EuroSAT (97.78% test accuracy) and its weights were frozen. The model was then deployed unmodified across CPU and GPU hardware at three numeric precisions (FP32, FP16, INT8), and accuracy, inference latency, throughput, and memory usage were measured for each of the five resulting configurations. Key findings: Accuracy is nearly invariant to platform and precision (97.70–97.73%, a spread of 0.03 points). Latency spans a 633-fold range, from 3.909 ms (GPU/FP32) to 2,473.969 ms (CPU/FP16). GPU inference is ~12.4× faster than CPU at matched FP32 precision. CPU-side INT8 quantization gives an 8.9× speedup over CPU/FP32 with no accuracy cost. CPU-side FP16 is ~51× slower with no benefit and should be avoided. Only two of five configurations are Pareto-optimal: GPU/FP32 and CPU/INT8. An independent replication on a second CPU (consumer laptop) confirms the same qualitative latency pattern, though its measured accuracy differs from the desktop — a discrepancy that is flagged and discussed rather than reconciled. Edge-device deployment (Raspberry Pi / Jetson-class hardware) was not measured directly, due to lack of access, and is reported as an explicit limitation rather than estimated. These measurements are used to propose a CNN Deployment Selection Framework that maps deployment constraint profiles (latency-critical, memory-constrained, accuracy-critical, balanced) to a recommended platform-precision configuration. This record archives the paper and the full experimental pipeline (training, precision export, benchmarking, and analysis code) for reproducibility.