When I originally used it at the beginning, if I recall, it was essentially a QEMU/KVM server with, indeed, a physical GPU passed to the VM, and each server had

When I originally used it at the beginning, if I recall, it was essentially a QEMU/KVM server with, indeed, a physical GPU passed to the VM, and each server had 2-4 GPUs assigned to it. When you logged in, you had to wait for one of the available servers to have an available GPU, in which point your virtual machine was mounted to that server and booted up, and their custom software on the VM would send the graphics back to you.