Virtualizing graphics processing in a provider network

A computer-implemented method, comprising:

• receiving, from a client by an elastic graphics service configured to provide virtual graphics processing units of a plurality of different virtual graphics processing unit GPU classes, a specification to provision a virtual compute instance, the specification comprising a selection of a virtual graphics processing unit (GPU) class from a plurality of different classes of virtual GPU classes, each class having different capabilities for graphics processing based at least in part on computational or memory resources, wherein the multi-tenant provider network comprises a plurality of computing devices configured to provide central processing unit (CPU) resources to implement a plurality of virtual compute instances that provide compute services over one or more networks to a plurality of distinct client devices of clients of the service provider;
• provisioning, for the client responsive to the receiving, the virtual compute instance from the multi-tenant provider network with a virtual graphics processing unit of the virtual GPU class selected by the client, the virtual GPU attached to the virtual compute instance and implemented on a physical GPU to provide graphics processing capabilities of the selected virtual GPU class to the virtual compute instance over a network, wherein the virtual GPU is implemented using a physical GPU that is located in a different computing device than a device of the plurality of computing devices that provides CPU resources to implement the virtual compute instance, and wherein the physical GPU is accessible to the virtual compute instance over a network;
• and
• executing, by the virtual compute instance, an application using, over the network, the virtual GPU attached to the provisioned virtual compute instance, wherein executing the application causes the virtual GPU to generate virtual GPU output that is transmitted over a network to a client device.

A non-transitory computer-readable storage medium storing program instructions computer-executable to perform:

• receiving, from a client of a plurality of clients of the compute services, a specification to provision a virtual compute instance, the specification comprising a selection of a virtual graphics processing unit (GPU) class from a plurality of different classes of virtual GPU classes, each class having different capabilities for graphics processing based at least in part on computational or memory resources;
• selecting a virtual compute instance implemented via central processing unit (CPU) resources of a physical compute instance of a multi-tenant provider network that provides compute services over one or more networks to a plurality of distinct client devices for clients of the multi-tenant provider network, wherein the virtual compute instance is selected based at least in part on the received specification and resources provided by the virtual compute instance;
• and
• provisioning, for the client responsive to the receiving, the virtual compute instance with a virtual GPU of the virtual GPU class selected by the client, the virtual GPU attached to the virtual compute instance over the network to provide graphics processing capabilities of the selected virtual GPU class to the virtual compute instance;
• wherein:
• the virtual GPU is accessible to the physical compute instance over a network,
• the virtual GPU is implemented via a physical graphics processor of the multi-tenant provider network, and
• the physical graphics processor is located in a different computing device than the physical computing device that provides the CPU resources to implement the virtual compute instance.