Use a GPU accelerator

You can configure worker hosts in your service-managed fleets to use one or more GPUs to accelerate processing your jobs. Using an accelerator can reduce the time that it takes to process a job, but can increase the cost of each worker instance. You should test your workloads to understand the trade offs between a fleet using GPU accelerators and fleets that don't.

GPUs are not available for fleets with wait-and-save intances.

Note

For testing purposes you are limited to one GPU. To request more for your production workloads, see Requesting a quota increase in the Service Quotas User Guide.

You decide whether your fleet will use GPU accelerators when you specify the worker instance capabilities. If you decide to use GPUs, you can specify the minimum and maximum number of GPUs for each instance, the types of GPU chips to use, and the runtime driver for the GPUs.

The available GPU accelerators are:

• T4 - NVIDIA T4 Tensor Core GPU

• A10G - NVIDIA A10G Tensor Core GPU

• L4 - NVIDIA L4 Tensor Core GPU

• L40s - NVIDIA L40S Tensor Core GPU

You can choose from the following runtime drivers:

• Latest - Use the latest runtime available for the chip. If you specify latest and a new version of the runtime is released, the new version of the runtime is used.

• grid:r570 - NVIDIA vGPU software 18

• grid:r535 - NVIDIA vGPU software 16

If you don't specify a runtime, Deadline Cloud uses latest as the default. However, if you have multiple accelerators and specify latest for some and leave others blank, Deadline Cloud raises an exception.