Well the way I understand this is that when you VMDP a device you are declaring it as a device that can be used directly from a Virtual Machine and therefore bypass the vmkernel for that device. It is like taking the device from the ESX server and having it ready for passthrough. At this point nothing really happens except of course that ESX server can no longer access the device, so be careful what you choose there as you may ruin your ESX installation. (for instance VMDP the sata controller where your ESX boot disk relies!!!)
You will also need to reboot the ESX server after configuring devices for passthrough.
So after that you will have to assign the VMDP device by editing the target VM's configuration and add the device found under "PCI devices". You can add up to 6 VMDP devices and of course you cannot have more than one VM's sharing the same VMDP device.
This will benefit you with performance but will sacrifice some of the nice features of ESXi hypervisor such as snapshots, hot plug disks or nics etc.
And of course will bound you with the need to reserve all the memory configured on the VM and the dreadfull pci.hole error when you try to give more than 2GB RAM to the VM.
It is a nice feature though this bug seems to spoil the fun.
On the screenshots attached you can see my VMDP configurations
