32GB RAM won't speed up single-GPU training.

Probably not. For training on a single GPU, the bottleneck is almost always the GPU’s VRAM, not your system RAM. More VRAM lets you fit bigger batches or larger models. More system RAM helps with data loading and preprocessing, but once the data is on the GPU, the RAM doesn’t affect training speed.

If you’re already running out of memory during data augmentation or shuffling huge datasets, then upgrading from 16GB to 32GB RAM might prevent swapping and keep your CPU feeding the GPU faster. But for most people training on a single GPU – even large models – 16GB is plenty. The money is better spent on a GPU with more VRAM.

Don’t buy 32GB RAM expecting faster training times. Buy it if your current setup is running out of memory during data prep. Otherwise, save your cash for the real bottleneck.