Estimate the GPU-minutes before you launch.
Pick what you want to do, the model, the GPU and the dataset. The estimator returns the GPU-hours the job is likely to take. The console prices that estimate with the current rate for your region.
Roughly how long it runs
An estimate for a typical run of this shape. Your data, sequence length and settings change it. Rates depend on GPU type and region; the console shows the priced estimate before you launch.
Questions, answered.
Why does the estimator not show a price?
Rates depend on GPU type, share and region and are shown in the console at launch time, so the estimator sticks to what it can predict: how many GPU-minutes the job is likely to take.
How accurate is the estimate?
It is a rule of thumb for a typical run of that shape and size; long sequences, unusual batch sizes and heavy evaluation loops can move it by a factor of two.
What is a GPU-minute?
One minute of one whole GPU; a pod on half a GPU for ten minutes uses five GPU-minutes, which is the unit QuickPods bill.
Can I stop early?
Yes, stopping the pod stops the meter, so an estimate is a plan, not a commitment.