Google announces the beta release of Graphical Processing Units (GPUs) attached to Preemptible Virtual Machines (VMs) in their cloud Platform. Google Cloud Platform (GCP) customers can now attach NVIDIA K80 and NVIDIA P100 GPUs to Preemptible VMs for respectively 0.22 and 0.73 dollar cent per GPU hour, 50 percent cheaper than GPUs connected to on-demand instances. This release is for Google a step further in providing their customers a choice of computational resources for high-throughput batch computing, machine learning, and scientific and technical workloads at a more granular level.
A Preemptible VM is an instance in which a GCP customer can create and run at significantly lower costs than standard on-demand instances. However, the Google Compute Engine can terminate (preempt) these instances after providing the customer a 30-second warning. Effectively, these instances can be used for a maximum of 24 hours. For GCP customers that have fault-tolerant workloads and do not require a dedicated instance, these Preemptible instances can be a good fit from a cost perspective. Furthermore, any GPU attached to a Preemptible VM is preemptible by default and thus billed lower.
Alex Hickey, an editor of CIO Dive’s website, provided some insights on leveraging Google’s Preemptiple GPUs in one of their latest briefs:
Building or running AI systems isn't cheap for the average company. With salaries for experts running well into the six-figure range and beyond, AI budgets can be hard to apportion. Hardware for computational processing is typically outsourced to save on costs. GPUs offer better speeds and processing times than dedicated hardware, which quickly piles on upfront and maintenance costs. Accessible tools, including processing hardware, are an important facet of AI and ML democratization. An estimated 40% of companies have AI pilots or experiments in place, and only around 20% have AI deployed at scale or in a core business function. But with more affordable GPUs, more companies may find room in budgets and strategies to get POCs and test cases off the ground.
A typical Preemptible VM is created by appending the --preemptible argument to the instance create command in the gcloud command line interface or when using the REST API the user sets the scheduling.preemptible property to true. Alternatively, the customer can set Preemptibility to "On" in the Google Cloud Platform Console and then attach a GPU as usual.
Image source: https://cloudplatform.googleblog.com/2018/01/introducing-preemptible-gpus-50-off.html
Furthermore, customers can have a dynamic pool of GPU power by creating a managed instance group with preemptible instances when needing more scalability. Note that in the instance template the preemptible option needs to be specified before creating the group. The benefit here is that preemptible instances, if sufficient capacity is available, can be re-created automatically when they're preempted. Currently, the preemptible GPU feature is available in the US-central1 region only. The full documentation on the Preemptible VMs is accessible through the Compute Engine documentation.
Google, Amazon, and Microsoft each offer computational resources, whether it is Preemptible VMs, spot- or reserved VM instances, at a low price. The differences lie in the flexibility of usage of the instances. Amazon EC2 Spot instances are comparable with Preemptible VM. However, customers cannot add GPUs to them. Reserved instances from AWS and Azure are available at a lost cost, yet come with a one-year or three-year period term. Depending on the use case and required availability, customers can either choose a shorter period with Preemptible VMs or AWS spot-instances, or a more extended period with Azure or AWS Reserved Instances. Both come at lower costs than on-demand instances on any cloud platform.