Energy-Efficient GPU Frequency Scaling Characterization for SLM Fine-Tuning on Embedded Platforms

Abstract

While embedded GPU dynamic voltage and frequency scaling (DVFS) is well-studied for inference workloads, fine-tuning exhibits different memory access patterns and runs 100–1000× longer, making inference-derived policies inappropriate. We present the first per-frequency characterization of transformer fine-tuning across three model scales (BERT-tiny 14M, BERT-base 110M, DeBERTa-xlarge 900M) on the NVIDIA Jetson AGX Orin, sweeping GPU frequencies from 306 to 1300 MHz on SST-2 and QNLI benchmarks. Across 77 experiments, optimal frequencies fall consistently in the 612–1020 MHz range, with production-scale models achieving 22–32% energy savings over the default governor. We develop a GPU-utilization-guided frequency selection algorithm requiring only 30 profiling steps that achieves a 1.5% average gap from the true optimum across 13 validation workloads, versus 21% energy waste for the default governor.