Separate resolution from batch size
A vision model that works on small images does not directly predict the footprint of your working images. First build a series at constant batch size with increasing resolutions, then a series at fixed resolution with increasing batch sizes. This lets you attribute the differences. Keep the same resizing and cropping rules: changing the preprocessing can shift the score as much as the choice of model.
Measure a full training or adaptation run
The 48 GB give more room than a 24 GB card for activations and auxiliary components. They do not, however, define a universally admissible model size. Measure backpropagation if you are training, the whole pipeline if you are producing embeddings, and the largest samples if shapes vary. Note the peak and the settings that caused it.
For adaptation, compare quality on a held-out set separate from your tuning settings. Increasing the batch size to fill all the memory is not, on its own, a scientific improvement.
Ada vs. Ampere: a comparison at constant recipes
The RTX 6000 Ada and the RTX A6000 both have 48 GB in this catalog, but belong to different architectures. So keep your CUDA, PyTorch and library versions as comparable as possible. Check compiled extensions for Ada and test the most specialized operation in your model. The L40S is another 48 GB comparison; an RTX 4090 may be enough if your peaks stay within its 24 GB envelope.
Prepare the data before the plan
Three days suit a resolution/batch mapping. Seven days let you adapt and evaluate a few variants. Thirty days support a vision campaign with repeats and organized exports. Prepare the image manifest, its splits and the evaluation script before choosing the duration. The order combines the configuration, the desired environment and your contact details; your crypto payment is then tracked in the same folder.