PUCC: a reproducible cloud-based framework for optimizing deep learning input pipelines and multi-GPU utilization

Abstract Cloud-based deep-learning systems can underuse allocated GPUs when storage access, preprocessing, and host-to-device transfer interrupt model execution. This study evaluates the Peak Utilization of Cloud Computing (PUCC) framework for improving node-level data delivery without changing the model or adding hardware. PUCC combines DUVRAM host-tensor staging with the HGPU+ execution pipeline. Four controlled conditions were tested: generator delivery (A0), staged tensors with generator delivery (A1), staged tensors with tf.data batching (A2), and the complete pipeline with AUTOTUNE prefetching (A3). The primary experiment used the same two NVIDIA Tesla T4 GPUs, DRDM1 split, EfficientNetB0–Vision Transformer, initialization, optimizer, global batch size, and five-epoch workload. Ten matched runs and external validation on a dual-T4 Colab Pro runtime, one Tesla T4, and a four-T4 multi-node system were completed. In the primary endpoint run, fit time decreased from 33.96 to 8.54 min, a 74.9% reduction and 3.98-fold speedup. Reconstructable throughput increased from 16.56 to 65.85 images/s, step latency decreased from 1523.3 to 732.0 ms, and GPU utilization increased from 45.3% to 97.9%. Peak VRAM per GPU increased from 4.15 to 11.05 GB. Ten-run and external analyses reproduced the performance direction. Accuracy remained a secondary outcome. PUCC improved input-bound training by reducing batch-delivery delays on allocated GPUs. The completed ablation, repeated-run analysis, and cross-platform validation support its reproducibility while defining memory and deployment boundaries.