Meningkatkan Efisiensi Energi Perangkat Edge melalui Optimasi Pruning dan Kuantisasi Model

Abstract

Edge computing devices are increasingly tasked with performing artificial intelligence inference under strict constraints on processing capacity and power consumption. This study evaluates magnitude-based weight pruning and dynamic quantization as practical model compression techniques for energy-efficient edge AI deployment. MobileNetV2, pretrained on ImageNet, was adapted to the CIFAR-10 classification task and compressed under three configurations: 40% L1 unstructured pruning followed by recovery fine-tuning (Prune40), dynamic INT8 post-training quantization (QuantINT8), and a sequential combination of both (Prune+Quant). All experiments were executed on a physical Intel N150 mini PC with a thermal design power of 6 watts, using PyTorch 2.1 in CPU-only inference mode. Results show that Prune40 reduced inference latency by 17.9% while simultaneously improving classification accuracy by 1.04 percentage points, attributed to the implicit regularisation effect of sparse weight removal and recovery fine-tuning. QuantINT8 yielded moderate latency savings (6.6%) with negligible accuracy loss. The combined pipeline achieved the lowest absolute latency at a marginal energy overhead. These findings establish magnitude pruning with recovery training as the most effective single-step compression strategy for low-power x86 edge platforms.