Every GPU we run, with the numbers that decide.

1. Does the model fit in memory?

   If the weights do not fit, nothing else matters — the job will not start, or it will spill to host memory and crawl. Count roughly 2 GB per billion parameters at FP16, 1 GB at FP8, 0.5 GB in 4-bit, then add about 20% for activations and the KV cache. Every model page does that arithmetic for you.

2. Is the memory fast enough?

   Token generation is bound by memory bandwidth, not by compute: the whole model is read from VRAM for every token. That is why an HBM card such as the H200 (4,800 GB/s) serves far faster than a GDDR card of the same capacity.

3. Do you need several GPUs to talk to each other?

   For training and for models split across cards, the link between GPUs becomes the bottleneck. NVLink models keep that traffic off the PCIe bus. For single-GPU inference it changes nothing, and you should take the cheaper card.

4. Then, and only then, compare prices

   Once two cards both fit and both serve fast enough, the decision is the hourly rate — and whether the job can survive a 2-minute notice, which is what moves it from on-demand to spot and cuts 30–53% off the bill.

Need a model we do not list, or more GPUs than a single node? Tell us what you need.