What GPU do I need, and what does it cost
What a workload demands is set by what you are running, not by which card tops a spec sheet. Each page states the rule it selects on, then lists every part that meets it with the lowest on-demand rate we hold.
One page per workload
GPUs for training a large language model
Datacentre parts with the memory to hold optimiser state, not just weights.
GPUs for fine-tuning with LoRA or QLoRA
One card, adapter weights only, and the base model quantised.
GPUs for serving a model
Enough memory to hold the weights, then throughput for what you pay.
GPUs for image generation
Diffusion models fit in far less than a language model.
The cheapest GPUs worth renting
The lowest hourly rate on record, with the memory to be useful.
The pages that are missing, and why
There is no page here for scientific computing, video transcoding or multi-GPU training. Those turn on FP64 throughput, hardware encoder blocks and interconnect bandwidth respectively, and we hold none of those figures. Writing them anyway would mean selecting on memory and calling it something else — which is the one thing a page titled “best GPUs for X” must not do.