Find the Bottleneck: Speed Up ML Pipelines

Modern GPUs keep getting faster each generation, enabling ever larger training and inference workloads. But in many real-world applications, GPU compute is no longer the slowest part: I/O and data movement are. This workshop teaches a practical, systematic approach to diagnosing and eliminating data-loading bottlenecks. Participants will learn how to design high-throughput data pipelines for both training and inference. By the end of the session, attendees will be equipped with techniques that typically yield 10% to 500% pipeline speedups — often without touching the model architecture at all.