1 article, newest first.
You bought a fast GPU and it's sitting at 30% utilisation. The model isn't the bottleneck — the input pipeline is. Here's how Dataset and DataLoader work, and how to stop starving the accelerator.