Google Cloud today announced the launch of preemptible GPUs. Like Google’s preemptible VMs (and AWS’s comparable spot instances), these GPUs come at a significant discount — in this case, 50 percent. But in return, Google may shut them down at any point if it needs these resources. All you get is a 30-second warning. You also can only use any given preemptible GPU for up to 24 hours.
This new feature applies to both the NVIDIA K80 and NVIDIA P100 GPUs that are currently available on the Google Cloud platform. Google will charge $0.22 per GPU hour for the K80 and $0.73 for the P100. In addition, you will need to pay for a VM. Google will only charge you preemptible VM prices for these, but you don’t have to worry about Google shutting down your VM before your GPU workload is done (or before you get your 30-second warning for the GPU shutdown).
Google notes that these preemptible GPUs should be a good fit for any fault-tolerant machine learning workloads and other computation-heavy workloads that run in batches. Using Google’s managed instance group feature, you also can automatically re-create your preemptible instance after it goes down.
Google’s move today comes only a few months after the company already significantly dropped its prices for non-preemptible GPUs.
In addition to this new pricing scheme, Google today noted that GPUs are now available in its cloud’s US-central1 region.