The right GPU depends entirely on what you are running. These pages give the sizing arithmetic and a hardware recommendation for each common workload.
From a single-GPU LoRA run to multi-node pre-training over InfiniBand.
Throughput-tuned nodes for serving models behind an API.
The best images-per-dollar on the fleet, on Ada-generation hardware.
Burst render capacity with hardware video encode.
Compute for a paper deadline, without a procurement cycle.
FP64-capable nodes for simulation, genomics and computational chemistry.