Google launched the beta version of a new lower-cost cloud computing option for people who want to throw the parallel processing capabilities of a GPU at tasks that can be interrupted. The feature, called Preemptible GPUs, now allows users to spin up instances in Google Cloud Platform using Nvidia chips that can be shut down at a moment’s notice.
Using GPUs in that way is half the price of spinning up a traditional on-demand instance in GCP, but it comes with a couple key drawbacks. Google can shut down the instance with only 30 seconds of warning, and preemptible instances can only run for a maximum of 24 hours. But even with those limitations, it can be a useful tool for developers with workloads that require massive amounts of raw computation and don’t need to maintain an internal state.
It’s also a win for Google, which would otherwise have this infrastructure sitting idle around its worldwide datacenter footprint. Google’s program is similar to Amazon Web Services’ Spot Instances offering, which also provides developers with access to discounted virtual machines that could be shut down with little warning.
Developers can deploy Nvidia K80 and P100 GPUs, which cost 22 cents and 73 cents per hour used, respectively. That cost comes alongside whatever they have to pay for the underlying compute instance, since Google makes it possible for people to tailor the CPU and RAM for all of their virtual machines.
Google also offers developers a way to automatically set up new preemptible virtual machines with GPUs in the event that theirs get shut down, using its managed instance groups feature. That will allow developers to define a set of instances that they want to run in Google’s cloud and automatically try to stand up new VMs to match the desired capacity.