Hi, everyone. Welcome back. In this lecture, I'm going to talk a little bit about graphics processing units, or GPUs. Google Compute Engine provides GPUs that you can add to VM instances. These GPUs are typically used to accelerate certain workloads that are run on compute instances. Workloads like data processing and machine learning are the types of workloads that benefit most from the use of GPUs. If you plan to create a compute instance that will leverage a GPU, you need to be sure that you choose which boot disk image that you're going to use for that instance. You also need to install the appropriate GPU driver on the instance. Some images like the Deep Learning VM images already have GPU drivers pre-installed, and they include packages like TensorFLow and PyTorch. Many public images, however, might require unique GPU drivers. Ultimately, you need to make sure that you identify what drivers you need for your images when leveraging GPUs. 

Before you attempt to attach a GPU to an instance, you should check your quotas page to ensure there are enough GPUs available within your project. You can, of course, always request a quota increase if there are not enough available. Whenever you provision an instance that includes a GPU, you need to ensure that you can configure the instance to terminate on host maintenance. This is because an instance with a GPU attached cannot live migrate. It cannot live migrate because the instance, under the covers, is assigned to specific hardware devices. So keep that in mind when working with GPUs. In the next lesson, I'm going to show you how to attach a GPU to an existing compute instance. I'll also show you how to install the necessary CUDA libraries to support the GPU.