The fastest way to pick a GPU on DeployGPU is to start from memory, not price: work out how many gigabytes a model or job needs, then let budget and interconnect narrow the field from there.
Start with memory
Every model and most training jobs have a hard memory floor before anything else matters. The models hub walks through that math for current open models, and three pages group DeployGPU's cards by the memory tier they cover: 48 GB, 80 GB and 96 GB and larger. Pick the smallest tier that comfortably fits the job, since memory that goes unused is still billed by the hour.
Then budget
Within a tier, the live prices in the table above are the next filter. A 48 GB job on an RTX A6000 usually costs less per hour than the same job spread across a pricier card, even when both fit it in memory. See how pricing works for what the hourly rate includes.
Then interconnect
If a job needs more than one GPU working together, either because the model is too large for one card or because training syncs gradients often, interconnect becomes the deciding factor. The multi-GPU page covers how NVLink and PCIe differ inside one VM, and how to check which one a specific offer has before launching it.
What is not on this list
RTX 4090 and RTX 5090 GPUs exist on DeployGPU only through Vast.ai, labelled testing only in the dashboard, and are not part of the families above.