Google LLC is introducing what it says is a more affordable way for organizations to run artificial intelligence projects in the public cloud.

The company today bolstered its infrastructure-as-a-service platform with the addition of Preemptible GPUs, which are on-demand graphics processing units for running machine learning workloads.

Companies can choose between Nvidia Corp.’s K80 and P100 chip models, which are widely used for machine learning because they can do lots of processing in parallel to crunch all the data needed to train and run the algorithms. According to Google, both are available for 50 percent less than the price of renting them as part of standard compute instances.

The reason is that the GPUs come attached to so-called Preemptible VMs, a type of low-cost virtual machine that the company originally introduced in 2015. The virtual machines, which are computers emulated in software for more flexibility and lower cost, run on unused hardware in Google’s cloud data centers. As a result, they cost considerably less than instances bought the usual way.

But there’s a catch. The availability of Preemptible VMs and Preemptible GPUs varies based on how a busy data center is. Consequently, Google only allows customers to rent them for a period of up to 24 hours. Moreover, the company may shut down an instance or GPU sooner if the infrastructure is needed elsewhere. The system provides only a 30-second notice in such scenarios so that applications have time to wind down safely.

As a result, Preemptible GPUs mainly appeal to workloads that don’t require continuous availability. They should be a particularly attractive option for handling the training phase of AI projects, which occurs before an application goes into production and tends to require a great deal of computational resources. Many companies would no doubt be willing to tolerate the occasional interruption in return for potentially significant cost savings.

The 50 percent discount carried by Preemptible GPUs comes on top of the 36 percent cut to graphics accelerator pricing that Google announced in November. Preemptible VMs are the company’s answer to the Spot Instances provided by cloud leader Amazon Web Services Inc., which likewise offers unused data center capacity at a discount but with a somewhat different approach.

Preemptible GPUs start at 22 cents per hour for an Nvidia K80 chip and 73 cents per hour for the Nvidia P100.

Image: Google