From 15% to 90% GPU Utilization: Fix the Data Pipeline, Not the Model
Fixing multimodal data pipelines boosted GPU utilization from 15% to 90% without touching the model.
“the hardest part of optimizing multimodal learning throughput isn't necessarily optimizing the kernels”
A practitioner walkthrough demonstrates that multimodal SFT pipelines routinely waste 85% of compute time on CPU-side data prep (loading, resizing, normalizing images/audio/video) before the GPU ever sees a token. Using Ray actors, prefetching, and zero-copy data transfer techniques on a Qwen-3-VL fine-tuning stack, the speaker raised GPU utilization from 15% to 90%. The insight is broadly applicable: for multimodal workloads, data engineering ROI often exceeds kernel-level optimization.