Accessing GPU inside a VM Bubble, Is It Possible?

[l008com]

I have a Mac Pro that runs High Sierra. It's CPUs are churning away running CPU work units of Folding and other projects.
But alas, none of them support GPUs on Macs, and this Mac has a decent GPU. To be specific, it has an OEM low end nVidia GPU, and an older mid-range AMD GPU, which I'm trying to sell and trade up to a modern mid-range GPU. The monitor is on the older nvidia card anyway. Is it possible for me to some how grand the Windows virtual machine full hardware access to the GPU so I can run GPU units in windows, while having everything else run on the Mac?
The machine has PLENTY of RAM so running two OS's shouldn't be an issue at all.
Also note, this is VMWare Fusion, their Mac VM client.

[l008com]

I could also use VirtualBox since it's free. If it can be done, it would have to be in one of those two options.

[ajm]

The Vbox doesn't support GPUs. You have to use a PCI pass through for that to work and as far as I know, only VSphere Hypervisor can do that. You can use it for free, but I finally never tried it.
https://www.vmware.com/products/vsphere-hypervisor.html

[PantherX]

In addition to what ajm mentioned, you would need 2 GPUs. 1 GPU for the Host system and 1 GPU for the VM to fold on. I am aware that Linux would be more efficient (higher PPD) with the current set of FahCore (the software that does the folding) but with AMD GPUs, it might be a bit tricky to get the GPU to fold under Linux. The initial steps would be to install the correct AMD drivers which has OpenCL support and then the OpenCL package (sudo apt-get install ocl-icd-opencl-dev) and then the FAHClient.

[ajm]

Linux is indeed significantly more efficient (than Windows) for CPU folding, but I haven't been able to demonstrate any consistent advantage when folding on GPUs, be it AMD's or Nvidia's.

[Joe_H]

Linux is indeed significantly more efficient (than Windows) for CPU folding, but I haven't been able to demonstrate any consistent advantage when folding on GPUs, be it AMD's or Nvidia's.

With the previous GPU folding core, Core_21, there was a definite linux advantage. Improvements in the Windows code that were included in Core_22 brought Windows and Linux GPU folding to basically even.

[MeeLee]

There's still an advantage on linux,
But the PPD difference isn't noticed much because of the less consistent PPD numbers per project on core 22, and due to the fact that Core 22 has much larger numbers.
If a 2080Ti gets 2.1M PPD in Windows, it gets 2.4M PPD on the same hardware in Linux.
On core 22 I've managed to get up to 4.4M PPD on the highest core 22 projects (stock cooler, OC and power capped), while most Windows users reported 3,6 to 4.1M PPD.
Just one user I remember, has reported 4.8M PPD on one WU, on a watercooled ROG STRIX, which by nature have the best binned GPUs.
The 0.4M PPD difference in PPD between his and my GPU, is mostly QRB. The actual difference in amount of work per hour would have probably been very little (1995Mhz on mine, vs 2050-2075Mhz on his).
But there's still a distinct difference between Linux and Windows.

[ajm]

Punctually, you can find anything and its contrary, but if you pay attention, if you note down all results during a period of time using Windows and then Linux, on the same kit, Linux doesn't produces more on GPUs.
https://docs.google.com/spreadsheets/d/ ... edit#gid=0

[l008com]

I do have two GPU's anyway so that's not a problem. I'll poke around with that hypervisor thing once the new card comes in.