AWS p3.16xlarge Windows instance reports only 2 GPUs available

I installed the latest SDK 9.1 with the included driver AND Visual Studio 13 CE on a Windows 2016 base AMI. I am able to run/debug my CUDA application, but the response to cudaGetDeviceCount is only 2 GPUs available.

There is supposed to be 8 GPUs on a p3.16xlarge instance. Is this a limitation of the current driver for Windows?

I assume that the correct number is reported for a Linux AMI since that seems more widely used. When can I use all the GPUs under Windows?

Turns out you have to install the display driver in the p3.16xlarge instance in order for the system & cuda to recognize all the GPUs, so it is best to skip driver installation if you’re installing in a different instance to create an image, and then install the driver in the instance type you want to use.