Spot instances
A spot instance is a GPU instance at a lower price per GPU-hour. In exchange, its GPU can be taken at any time when the capacity is needed, with about 5 seconds' notice. Use spot for work you can restart: training that saves checkpoints, batch jobs, experiments.
Spot is offered where your launch dialog shows it. If you don't see the choice, it isn't on sale for that GPU or region right now.
Launching a spot instance
- Go to Compute → GPU Instances and click Launch instance.
- Under Pricing, pick Spot instead of On-demand. Both show their price per GPU-hour.
- Choose a duration (see below), fill in the rest of the form as for any instance, and click Launch.
A spot instance has at most 1, 2 or 4 GPUs (the dialog says which), and your account can run a limited number of spot instances at once. If no spot GPU is free for the size you chose, the dialog says so: try fewer GPUs, try later, or use on-demand.
What happens when the capacity is needed
It depends on the duration you chose.
With a fixed duration (1–72 hours), it pauses and comes back on its own.
- Its status shows Paused (spot), and it is not billed while paused.
- It resumes automatically when a GPU frees up, restarted from its image, with its drives and filesystems attached.
- It runs until its running time reaches the duration you chose, however many times it is paused. Time spent paused is added to its end time.
- You get a notification each time it pauses and resumes.
- You can terminate it at any time. You are billed only for the time it ran, and the rest of the hold is refunded.
With no end time, it stops and does not come back. You get a notification, you are billed only until it stopped, and the rest of the hold is refunded. Choose a duration if you want it to pause and resume instead.
Either way, memory, running processes and the container's own disk are lost. Save your work to a cloud drive or shared filesystem, and make your job able to start again from its last checkpoint.
What spot isn't available for
Spot is for single GPU instances. It isn't offered for GPU clusters, model instances or reserved nodes.
From the API
Launch with "interruptible": true where your availability shows a spot price. The fields, limits and statuses are in GPU instances API → Spot instances.