As previously announced, Google is now making the preemptible pricing option available for instances running on GPUs. While the company originally said the option would reduce prices about 50 percent, the actual discount is more like 70 percent. For example, high-end Nvidia Tesla V100 GPUs run $0.74 per GPU hour on the preemptible plan, as opposed to $2.48 per hour on the on-demand plan.
“Preemptible GPUs are ideal for customers with short-lived, fault-tolerant and batch workloads such as machine learning (ML) and high-performance computing (HPC),” blogged Google Product Manager Chris Kleban.