• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

Discussion AMD ZeroCore KVM GPU passthrough tests

PingSpike

Lifer
Feb 25, 2004
21,234
183
106
Well, awhile back I read about AMD Zerocore, where with the display off the GPU with power down and supposedly use ~1 watt...or zero watts...or 3 watts. Depends who you ask. Anyway, here's some quick info about it:

Introduced with HD7000 series cards back in 2012 or thereabouts.
Only works with VGA or DVI connections, not HDMI.
Forum posts seem to indicate that it has never and likely will never work on Windows 10.
The GPU should be mostly shut off in this state with juice only used to keep the PCIe bus active. Or at least that is how it was sometimes described.

Techpowerup mentions it using 1.11watts in this state, measured at the graphics card itself after the powersupply. They mention this same number in the review of several cards which makes me think they probably only tested it once and copy and pasted that blurb. That's still better than other review sites that seemed to just copy and past AMDs marketing materials.

I was excited when I read about this, because it meant I could have a GPU passthrough virtual machine that used very close to no extra power when not in use. One of the disadvantages of GPU passthrough in my main server has been the card just sits there most of the time using 10watts even when the VM is shut off. I also already own a MSI R7 250 DDR3 and an MSI R7 360 to test with.

Testing with a kill-a-watt and subtracting the wattage used by a system with the card not installed I found that:
250DDR3:
CardIdle VM Not StartedIdle ZeroCore Active Running VMDriver
R7 250 DDR3~7watt~6wattCatalyst 14.4
R7 360~8watt~7wattCatalyst 15.7.1

So that was a pretty disappointing finding. The results are kind of noisy, but the 1watt difference seemed noteworthy. The fan does shutoff on the card, but that seems like it probably accounts for all the power savings versus just shutting it off and letting linux host downclock it. Perhaps this feature is broken on the bios of the cards or maybe it never really worked to begin with. Given that the Windows VM seems to have spikes of activity that increase presumably CPU and disk power consumption while it is running I actually think turning the fan off and running Windows probably wastes more energy. There is also the possibility that it would work differently on bare metal. That doesn't help my use case, and given it doesn't seem to work on Windows 10 at all (I haven't heard of anyone getting it to work on linux period) it probably doesn't help anyone else either.

I don't think I'm willing to purchase additional cards to test this but its possible some cards work better than others.

Any ideas? My bios has ASPM options disabled, I've wondered if those could be a factor. But I'm not sure messing with them will leave everything else working.
 

ASK THE COMMUNITY