7 Years Late is Better than Never.

[Emily Young]

For years, Nvidia has tried to artificially block GPU passthrough on GeForce cards, but now they’ve finally made a 180. Could this be the start of a new Nvidia?
 

 

Buy NVIDIA RTX 3080 GPU (PAID LINK): https://geni.us/Si8kya

Buy AMD Radeon RX 6800 XT GPU (PAID LINK): https://geni.us/UsUD

Buy a Ryzen CPU (PAID LINK): https://geni.us/SOqq

Buy an Intel CPU (PAID LINK): https://geni.us/AM9ILRm

Purchases made through some store links may provide some compensation to Linus Media Group.

[Quackers101]

youtube so weird, when it's just released you might not be able to view it but also able to view it.

And sometimes you can access it through other means (youtube lists or different pages).

but it's up now after 12m? derp

 

[JoSiau]

I couldn't say based on the video if it is feasible today to have gpu pass-through on vm using any virtualbox or else? I recently researched about that subject but there are only paid solutions that allow gpu pass-through right?

[Radium_Angel]

6 minutes ago, JoSiau said:

I couldn't say based on the video if it is feasible today to have gpu pass-through on vm using any virtualbox or else? I recently researched about that subject but there are only paid solutions that allow gpu pass-through right?

IME that is correct, you need the right drivers, the right video card, and the right VM software to make it all work.

All of that means money, lots of money, as it used to be only Quadros and the paid-for version of VMWare, etc.

And it was *still* a PITA to set up.

 

[arcturus01]

I have herd lots about this happening; tho not what it means for someone who already has gotten past error43

 

I have KVM running and a windows vm. running an nvidia gpu. it works. do i gain any features from this change?

 

 

[Hunter259]

GPU Partitioning (SR-IOV sorta kinda) exists in Hyper-V as of a few Windows releases. Unsure if it's on Win server or base windows pro but it's SUPER interesting as a workaround.

[Nystemy]

The main reason I can think of why they did this were likely so that all data centers wouldn't just run out buying cheap GTX cards for their VMs instead of Quadros.
After all, there is a lot of professional applications that doesn't need any of the advantages of a Quadro, but the advantage of a VM and PCIe pass through is fairly huge in itself. Since moving a MV to new hardware is a lot easier than moving a bare metal machine to new hardware. And considering the price difference between the two, I am not surprised that nVidia made the decision they did back then.

All though, Quadro cards are technically more than just a driver difference. They primarily have ECC memory, but also IIRC, they are validated to have bellow a certain error rate in their calculations. And yes, calculation errors can be fairly major depending on the application, for larger computing tasks the errors will add up and make the end result inaccurate. A lower error rate makes the end result more accurate. This is however not always important, so a DC could see reason to just use a GTX card instead. And that is money that nVidia then doesn't make.

 

And why one wouldn't just run and grab an AMD card that supports PCIe pass-through is down to the application, if the application is QUDA accelerated then an AMD GPU is fairly pointless.

 

And why one would want to move a VM, typically service/maintenance related reasons. Not to mention software licenses and various DRM schemes. Easier to tie the license to a VM and move that, than tie it to a bare metal system. One could move the license, but that is a hassle compared to moving a VM, especially if the machine has multiple licenses.

Why nVidia has now decided to support PCIe pass-through on their consumer cards is a good question.
Here is some reasons I think nVidia might be releasing this feature.

• They might have an increased competition from AMD in the datacenter.
• The new 3000 series is potentially still having low yield. This can impact error rates, even if a lot of yield loss is due to contamination or misalignment of features. (It is practically the silicone lottery.)
• Or maybe there is just too much demand in the DC space to support the current work from home migration requiring more cloud infrastructure.
• Or consumers are buying quadro cards since RTX cards are rare these days.

nVidia might currently be forced to open up all cards for pass-through to satisfy their datacenter customers (a fair portion of their market that also ends up affecting everyone else relying on cloud services), since nVidia can't simply take any GPU chip and make it a Quadro. And making a "datacenter" skew that only has PCIe pass-through as the highlight feature would raise criticism, especially if it is having a noticeably higher price than the equivalent RTX card, a datacenter operator would just think nVidia is silly at that point.

[Egg-Roll]

Fun fact, car manufactures can void your warranty if you use your vehicle out of the its designated/designed use, so in a way yes they could force you to buy a commercial fleet vehicle (or a truck). Equally so can your insurance provider cause you grief (read the ToS, it should say something like not to be used for commercial purposes). While Linus is correct no one can "force" you, just don't expect your warranty claims to go threw or your insurance to cover any damage on a "business run" less it being included in your policy (unlikely).

 

2 hours ago, GabenJr said:

For years, Nvidia has tried to artificially block GPU passthrough on GeForce cards, but now they’ve finally made a 180. Could this be the start of a new Nvidia

Serious question, why did the editor feel the need to add a engine noise to a EV?

[Bitter]

So I still can't pass my AMD card through VirtualBox on a Win10Pro host to Linux/Win10 guests?

[Emily Young]

2 hours ago, JoSiau said:

I couldn't say based on the video if it is feasible today to have gpu pass-through on vm using any virtualbox or else? I recently researched about that subject but there are only paid solutions that allow gpu pass-through right?

9 minutes ago, Bitter said:

So I still can't pass my AMD card through VirtualBox on a Win10Pro host to Linux/Win10 guests?

VirtualBox doesn't have this feature, no, but Linux KVM / vfio does and is also free. This is what we used for demonstration purposes in the video.

 

2 hours ago, arcturus01 said:

I have herd lots about this happening; tho not what it means for someone who already has gotten past error43

 

I have KVM running and a windows vm. running an nvidia gpu. it works. do i gain any features from this change?

If you're already up and running, then not much. You can remove the workarounds and have a "pure" VM, and you won't have to worry about something breaking due to changes in Nvidia's VM detection or what have you. It just makes your life easier, or more accurately, it makes your life as easy as it should have been in the first place.

 

41 minutes ago, Egg-Roll said:

Serious question, why did the editor feel the need to add a engine noise to a EV?

Either they felt it was "weird" to have no audio, they weren't aware the Kona was an EV and audio wasn't rolling when they shot that scene, or they wanted to see how many people would notice.

[das1996]

What is the hypervisor in use at 71s - https://youtu.be/JGvrXXonoqM?t=71

 

Never mind, I see it answered above, Linux KVM / vfio .

[Egg-Roll]

15 minutes ago, GabenJr said:

Either they felt it was "weird" to have no audio, they weren't aware the Kona was an EV and audio wasn't rolling when they shot that scene, or they wanted to see how many people would notice.

Fair enough, the Kona did start off as a normal gas vehicle tho not sure if all formats got introduced the same year/time here or not.

Run it past Linus (and owner) if we could get it on Short Circuit or something not enough people talk about EVs. Plus it's one of the few vehicles you can actually get for around 45K CAD with actual range

[Bitter]

44 minutes ago, GabenJr said:

VirtualBox doesn't have this feature, no, but Linux KVM / vfio does and is also free. This is what we used for demonstration purposes in the video.

I'll be sure to check that out, thanks. I can dream that maybe someday I can pass from Win10 host to guests, that would be nice. Until then I'll have to setup dual boot and learn some things, thank you!

[macmanluke]

Does anyone know when this changed? was it a particular driver?
Iv been passing through a 3060 in unraid to windows no problems for over a month now but maybe it was already fixed?

[Luscious]

I remember this question cropped up for me on the F@H forum a few years back. Linux always gave higher PPD than Windows, but unless you had GRID compatible cards or ran thru hyper visor it was impossible to get GPU pass thru to work. Yeah, if you had Linux installed as a base and Windows running in a VM it wouldn't be a problem, but going the other way around, absolutely no chance.

 

Because yes, if the limitation is removed, it would now allow F@H the ability to work in Linux via VM on consumer cards. That's good news for the folding community!

 

Maybe nVidia can;t get enough GRID compatible cards out there for those who need it? Maybe not everybody is running Windows Server OS. It's clear something or some group has compelled nVidia to change their mind.

 

Could very well be nVidia has their hand in the problem and not just the solution. Consider those mining right now who for some reason need to go thru a VM - they won't be buying GRID cards or coughing up dough for a server license. Is it possible MONEY is the motivator behind the decision and nVidia wants their slice of that cake again? Just how greedy is that guy in the the leather jacket?

 

There's also that 32 GPU head limitation of Windows - running them in a VM on Linux and passing the cards through to Windows would get around that issue as well. That's especially important if your 4U box happens to hold 10 or even 20 cards and you don't want to split resources in two.

 

Gotta question as well if this decision hinders or helps the current shortage, because if this just takes more consumer cards AWAY from gamers and puts them to use in VM farms, that's another big blow.

[sebbe66]

I can't semm to finde the new  pass-through driver on Nvideas webbsite.

Is it releast yeet ?

[das1996]

https://forums.servethehome.com/index.php?threads/official-geforce-passthrough-coming.32443/

 

Was posted back on 4/1/2021 on serve the home.  I'm still running an older driver with a work around (disable/enable gpu on shutdown/restart).  I'm actually thinking of moving the gpu to a dedicated htpc rather than the esxi box. Need to determine if there will be any additional power savings this way. 

 

I'd like to think idling the gpu 22 out of 24 hours a day uses more energy than a dedicated htpc for 2 hours.

[dmcf]

Likely because VGPU_Unlock was recently released.

[TobyCat]

I'm very excited about this in terms of Linux usage.  It's been a while now where I've wanted to make the switch from Windows 10 to Linux, but the only thing keeping me away not knowing if my single 1660 Super will work out for GPU passthough, and i don't want to go through all the crazy workarounds until i know i have the processing power.

 

So now, if it really is that much easier, i can just try it and see if i need to amp up my GPU.

[Lucas Zanella]

I just tried it in a brand new Windows VM, updated the Windows, tried to install the NVIDIA geforce experience driver, and still get code 43. Anyone knows what to do? It's a GTX 1660, Ubuntu 20.04 host. It worked before with the usual workarounds

[das1996]

^^Tried it in an existing windows vm.. gtx 1050ti, esxi host.  Error 43.  So maybe there's other things that need to be adjusted.  I reverted back to a previous snapshot.  Had to reboot the server because the video card was left in a poor state.

 

I'm about ready to just use a dedicated box for htpc instead of gpu passthrough.

 

I'm curious how much power this gtx 1050ti consumes while the windows vm is shutdown or sleeping.  I wonder if the power usage of a dedicated htpc box will be less than that of the esxi box with gpu.

[Quackers101]

13 hours ago, das1996 said:

^^Tried it in an existing windows vm.. gtx 1050ti, esxi host.  Error 43.  So maybe there's other things that need to be adjusted.  I reverted back to a previous snapshot.  Had to reboot the server because the video card was left in a poor state.

*finds out only the newer cards have the update and older ones might get it*

Although unsure, but sounds like nvidia if it is, can be wrong though.

[igormp]

On 4/25/2021 at 6:07 PM, Bitter said:

I can dream that maybe someday I can pass from Win10 host to guests

It's possible to do so on Hyper-V. You can even get GPU partitioning (GPU-P), as mentioned by someone above.

[Bitter]

4 hours ago, igormp said:

It's possible to do so on Hyper-V. You can even get GPU partitioning (GPU-P), as mentioned by someone above.

I'll also look at that, thank you.

[Bitter]

On 5/4/2021 at 8:31 AM, igormp said:

It's possible to do so on Hyper-V. You can even get GPU partitioning (GPU-P), as mentioned by someone above.

Finally sat down to work on this, have the VM made, now I need some ELI5 style directions or a link to some directions dumb enough for a mechanic to pretend to be an IT guy for a few hours to get it running. Help/suggestions appreciated.