Use ANY Headless GPU for Gaming in a Virtual Machine!
HTML-код
- Опубликовано: 2 сен 2021
- HUGE Thanks to Linode for sponsoring today's video. Get a $100 60-Day credit just for signing up for a new account at linode.com/CraftComputing.
For YEARS I've been trying to make this work. Now, you can use ANY graphics card for 3D Acceleration in a virtual machine. This means any nVidia Tesla or AMD FirePro card normally reserved for machine learning can now be used by you to play games.
But first... What am I drinking???
From Treehouse Brewing comes the Perfect Storm American Double IPA @8.0%. What starts as a mandarin orange juice bomb quickly turns into 'every other hazy IPA' I've ever had. Get a tulip, not a pint.
Full written instructions + software downloads: drive.google.com/drive/folder...
Links to items below may be affiliate links for which I may be compensated
Tesla K80 for $170 on eBay: ebay.to/3iAHkX9
Tesla M40: ebay.to/3DlPPxu
MERCH! GET YOUR MERCH HERE! Shirts, Hoodies, Glassware!
craftcomputing.store
Follow me on Twitter @CraftComputing
Support me on Patreon or Floatplane and get access to my exclusive Discord server. Chat with myself and the other hosts on Talking Heads all week long.
/ craftcomputing
www.floatplane.com/channel/Cr...
Music:
No Good Layabout by Kevin MacLeod
Link: incompetech.filmmusic.io/song...
License: filmmusic.io/standard-license 
Наука
This is like the peak of the mountain of troubleshooting!
I can't tell you how satisfying it is to see such a clean and functional set up.
"but this is craft computing annnd I'm Jeff" 😂 😂 😂
I know I'm about, son.
@@CraftComputing right at the "but" I thought "it'd be funny as hell if he... Omg he did it!". I laughed so hard I choked on what beer didn't go flying
The man, not a myth, but a legend
"Do not check the primary GPU..."
MY GOD! I have been struggling to get my GRID K1 and Tesla M40s working and just couldn't figure it out in PVE5, 6 and 7. Now I can finally ditch Server 2012 for the host OS.
YOU ARE AWESOME!!!!
I know I'll never be doing any of this but the nerd in me loves these videos!
Same here. I do an ebay search every time even though I have a Rx 6800XT in my gaming PC.
Yeah, actual facts
Tesla cards aside, Jeff is without a doubt stepping up his beer game!
I will never do any of this, but it's so interesting to see you tinker :) You and Level1Techs are doing to great work in these niche areas of PC land.
Still contemplating about setting up a proper home lab, but for the time being, I just watch and learn from your shenanigans, and I just love it!
It seems you've been pretty busy on your conquest
Really enjoying your content !
I'm going to try this setup. Thanks for the video, I've been looking into how to do this for a while now
You are really making me prioritize building my server/workshop room before any other home improvement project. My NAS will not be happy sitting next to my desk for too much longer. Now I need get my v-machine box up and running....
Glad I could help you with the usbmmidd driver. :D
It was a putting 2+2+2 together with everything else that made this happen :-)
I love theese videos!! You make it seem so easy, but i'm guessing it takes alot of work to get to that point.
Literally been looking for this method for years.
very cool. this is the exact sort of solution i was just looking into so my kids could play real pc games. honestly been thinking of setting my whole house up like this would save a lot of space we have 8 machines running here and i would love to consolidate that all into one rack.
This couldn't have come at a better time... I was just looking at using K80s (plus an older Threadripper) to run VMs for gaming for the family. I was unaware of the requirement that one VM of the two to not be running. Thanks so much for the heads up! It definitely decreases the interest level in using them (except for in the VM I'm going to be running for video streaming and gaming at once, but may look at different options for everyone else). I'm also going to second the community-driven FOSS Moonlight program. I find it to be better than Parsec at home and on par or slightly better over the internet.
Thank you Jeff, Great video as always
I was just starting to mess with my virtualisation monster when I got the notification for this video: windows server running on Unraid aiming to get GPU-P working so I could get rid of my dedicated Hyper-V machine. I listed my two M40’s on eBay yesterday because I didn’t have the time to figure out the display issues, this gives me fresh hope.
Wish me luck!
Love it! Next phase, seeing how lean can you go on the client device. Thin client? Zero client? RPi?? Probably just minimum or recommended system requirements for Parsec I would think...
It is midnight but a new video is a new video
This video has answered so many questions thank you.
Thank you so much for this amazing guide!!
Excellent find. Thank you.
The video I've been waiting for. Absolutely amazing work Jeff! I just wish the K80 could be cooled better for a in-room home lob.
Yeah, it's one VERY deep heatsink, and requires a ton of pressure.
I’m aiming to put an NZXT kraken G12 on at least one of my M40’s using an older Corsair AIO I have. It’s a fairly simple bracket so it’s probably easy to adapt or 3D print something that’ll add two AIO’s to the K80
There are full cover gpu blocks for the k80 if I remember correctly
How about a mineral oil dunk
I pass through an old XFX 7550 Ghost single slot to a VM. It's a fun toy and Mesa drivers are so much better these days.
Lot of work to make this video. Great job sir
This Channel made me switch from desktop hardware to enterprise system. So cheap for so much performance with raid dual and quad socket cpu configurations and more
Thank you, I was able to get more performance from my K10 converted to Grid K2. I also got the K80 working and it's not great but not Bad.
This is pretty cool. Looks pretty good for running remotely on a VM, on a Proxmox server
I have an Nvidia tesla M40 in my machine. With a 980ti heatsink and fans on it for cooling. A little modding was needed but works great! the Pcb's are almost identical.
Wow, what a journey, let's do this ! Going to add Garuda gaming distro as well.
So I just followed this tutorial(though I'm using manjaro and a tesla m40) and IT WORKS! One issue, just like you, is cooling I have a 3d printed adapter for a server fan and at 12V it's deafeningly loud 5V is just fine for minor games but it just isn't enough for most the games I'm trying to play. So, my plan is to modify a dead RX 480 blower that is much quieter even at 12V and I would be able to run the fan on a motherboard header instead. Hope this works!
I’m so irritated I didn’t find this channel sooner been trying to do this kind of stuff and looking for tutorials for so long and now I have them thnx
Magnificent work :)
Thank you for making this video. I hope to use steam and a steam link device. Wish me luck!
I do wonder if the old arctic cooling kit for the GTX690 would fit the K80...it was made with two discrete coolers for each gpu chip and used stick on ram sinks.
No way. I was just wondering about this earlier today. Look at Twitter, I tagged you in a tweet about this earlier 😂
Could this work with a Tesla P4? Thank you
Wow this is cool maybe one of my future hobby project
0:12 What a beautiful cat❤
Hi, could this work with ESXI 7 or will you still be limited by being forced to use Nvidia Grid to allocate the 2 GPUs on the K80?
Awesome method! The way to survive scalperpocalypse!
May try this for gaming lounge
This methode works to for an rtx 2060 👌🏼 thanks man 😁
What kind of server did you put these cards in and do you think a setup where you put several of these cards in would make for a viable daily driver PC or would you start to worry about PCIe-lanes running out or becoming CPU bound?
I have to try this now.
I've been needing this. thank you! would this work with quadro cards as well?
I tried partitioning a Titan X Pascal inside a Windows VM using Hyper V for nested virtualization but it performed so bad its unusable. I'm guessing nested virtualization cause a lot of io wait and therefore laggyness. Was testing on an i7 5820K + Asus X99 Deluxe U3.1.
Spacedesk looks to be the answer to the problems getting an output from my Tesla m40 on my guests. Parsec worked fine on the host OS as described here but couldn’t grab the screen on the VMs.
My windows 10 hyper-v box sits on proxmox so I have nested virtualisation happening too, this isn’t wholly crazy as I think it’s the most effective way to have two graphics cards working with paravirtualisation.
Now I have an excuse to buy a thread ripper and less guilt towards the imaginary funds set aside for three computers for the children.
Thanks for setting me on this path Jeff!
Thx. For the update.
This is really an interesting setup,
Can you elaborate the network issues with Linux VMs? I have an ubuntu vm running on proxmox with PCIs passthrough without any problems? (Or I just don't see them)
There are so many little clever tricks in this video it's admirable. Using a virtual display driver to force vnc to use another display driver so when noVNC shows up a blank screen the primary display adapter will be pushed through tightVNCs server. Then using the policy editor to force the screen on at each boot. THEN streaming a DirectX devices draw calls remotely with couch play software to fill in the gaps. All while benchmarking the second GPU on the same PCB in an adjacent VM, love it.
you seem to know alot about servers lol, why wont the virutal display driver run on my machine? The driver has a error 31
@@II_superluminal_II error 31 doesn't mean much without a lot of context. A few things suggest it's a permissions issue like the installer wasn't ran as administrator. It depends on what executable returned the error so that the error number corresponds to the appropriate lookup table.
Thanks for the interesting information. Can this method be used to improve video encoding performance? DaVinci Resolve 17.
Great Video, today i ordert a Nvidia K80 for my Dell PowerEdge R730. Hopefully the graphics card will run as well as your solution under VMWare ESXI 7
Great stuff. By the way, I use your glass of beer as the progress bar :)
This guide is not only useful with Proxmox on a server passing a Tesla M40 to Windows VM (which your information here helped me set up, works fine), I also dualboot to just Windows 10 on the baremetal Poweredge R720 server for testing, and your setup with Parsec and the virtual display driver (I used usbmmidd_v2 instead of V1), is also useful there!
For example, on bare-metal Windows 10, if I use Parsec to just watch the 'hardware' Matrox built in graphics on the server, at 1080p (with the registry adjusted to unlock 1080p on that ancient Matrox built in card), I get 30ms encode in Parsec (not great)
But if I use Parsec to instead watch the virtual fake display like you have setup here, the encode actually kicks in on the Tesla GPU and the Encode latency time is down to 1-3ms or so (10x better).
Basically tl;dr your guide here is helpful not just for Proxmox but also even for just running bare metal Windows 10 on one of these headless servers and trying to get good latency in Cyperpunk 2077 remotely using a Tesla M40.
Exactly the video I need, I'm looking for a GPU to use with VMs. Need Windows for college, and there's a course with OpenGL, that should go much better with my VM able to use a whole GPU.
I don't have a huge amount of PSU capacity headroom, though while it's idle(which it would be when I'm not hammering it in my VM, and I'll mostly be not be hammering my main GPU at the same time) it should be OK.
Being able to use enterprise cards might open up options, though I'm currently leaning to an RX6400 for cost reasons.
If the only intended purpose of a specific virtual machine is gaming, would it make sense to have it set to automatically boot into steam big picture, so that even something like an ayn odin could use steam link to connect? I know that it would barely be any different from setting up a dedicated physical machine, but I imagine the virtualization of such a menial task would be reasonable for the space savings, especially because multiple virtual machines could be set up for other functions without the added use of space.
The madman did it!
Nice! Have you tried this with a Linux VM with Steam or Lutris?
OHHH This need more IT knowledge that i thought :D
Hi great video !!!! in your opinion tesla M40 to be used with proxmox windows vm does need Above4G decode enabled?? my mainboard don'have this feature. Many thanks!
Could I run one of this cards in Ubuntu, setting a passthrogh for a windows VM with VMmanager and visualizing it with Looking glass?. It runs well with my gtx 950 without plugging it to my monitor or a dummy plug. great videos!!!!!!
Parsec is amazing I use it for my Windows and Bigsur vm's on my Unraid server. Great video btw
Bigsur VM’s? On Unraid
Oh Parsec released support for MacOS fairly recently it seems, crazy. Thnx for the heads up
16:22 parsec only supports the raspberry pi 3, not anything else. Since they have to access the encoder directly instead of through an api, it's much more work to develop an app for all rpi devices
I use usbmmidd to run parsec in a optimus, no mux switch laptop with the lid closed, so when you said that there was a solution, I bet that usbmmidd was involved.
why not remove the shroud and strap fans directly to the heatsink? or is the heatsink part of the shroud on these/ not dense enough for regular fans. I have 2 120mm fans on an rx480 and it runs cool and quiet. I realise these are higher thermal output (and they look pretty with the shrouds) but is this not an option?
It is cool but it doesn't support gpu virtualisation so you are limited to the count of gpu's you have. PCI passthrough is enabled on 30 series too. So you can also use them. They don't have to be headless, actually if they have hdmi ports it is better. Just plug them into monitors, passthrough usb ports to add keyboard, mouse etc to the virtual machines, then viola, you have multiple gaming computers running on a single server.
Also using hdmi ports will remove the parsec's blurry gaming experience. It will be just like a physical computer. No additional configuration is needed! Just plug it and it works!
Yet, a lot of games won't support virtual machines especially those with anti-cheat systems. They don't like to be run in virtual machines. Including CS:GO, valorant, LoL etc.
Some games will just crash, in that case change you cpu type to host and enable Intel VMX or AMD SVM!
Some games will really suffer from limited ram performance. I could hardly get 40fps on cyberpunk on my 3090 (normally it would get around 60-70).
Also enabling pci passthrough will disable memory balooning. Which makes the virtual machines use the full amount of memory they have.
So if you have headless gpus, I HIGHLY recommend using gpu-p on Hyper-V.
This allows you to use memory balooning system of Hyper-V.
However with this method, you can run multiple cryptocurrency mining OSes on each gpu flawlessly. I highly recommend proxmox for that particular use case.
When parsec program is closed on client side does virtual monitor work 1080p in background not returning to 1024x768? so far i'm using teamviewer it works sometimes. I want to use camera stream 24/7 via OBS it has to go through the camera app without rtsp url
Oué mec tout marche comme presenté je te kiff
I wonder if this will work just as effective as Steam’s own “Steam Link”, if you pass a gpu via a VM and launch Steam itself inside of it.
You could use Steam link instead of Parsec for streaming, but the USBMMIDD setup would still be required to emulate a monitor.
@@CraftComputing for a couple months when I had no monitor for my 1060, I used Steam Link to use my (now dead) MacBook Pro as a monitor for gaming, and I had no video output connected at all. No headless adapter, no usbmidd, none of the ports had anything connected at all. Parsec on the other hand, did not work, because it needs a video output device or monitor.
Fyi...
Depending on which pve you're using or if you're using grub or systemd boot... zfs on root may also affect...
But instead of /etc/default/grub
You may have to edit /etc/kernel/cmdline
And instead of update-grub
You may have to issue proxmox-boot-tool refresh OR pve-efiboot-tool refresh
Depends on the version of pve -
Depends if grub or systemd was the BL -
Depends if using UEFI
Unsure which is affecting the process - however if your wondering why your passthrough is not working follow what Jeff is doing - if it didn't work - try the other variations - one will work.
Hopefully I can get a response, I've done everything in the tutorial, but I'm still met with a black screen in parsec. Remoting in with RDP or VNC shows encoding activity for the gpu when I attempt to connect with parsec, but still black display. Generic PNP monitor is selected, and I've disabled the VGA display from ESXi
Thanks Jeff!!!!
Hi Jeff. I have a Tesla k80 with 24GB of ddr5 ram. And I’m running esxi. Do you think I might be able to replicate your setup with this GPU. I want to use it for some data science and ml training and of course gaming. But gaming is a must
I wonder if you can comment on the following question? If I run headless GPUs inside a GPU server (Gigabyte) to do Compute (BOINC/F@H), and I then access Ubuntu desktop via IPMI (iKVM) will it work? There are some reports that you just get a blank screen as the BMC display cannot output if dedicated GPUs are installed; although these are mainly about Supermicro servers. I have searched Google high and low but cannot seem to find a definite answer on this. Your opinion on this would be most helpful.
Jeff, do you know if i could passthrough the primary gpu (my only gpu)? Or will this only work for multi-gpu setup? My system doesnt have a built-in video card and i have a headless gpu.
What cpu or cpus were you running? and what allocation of cores for each machine? im running 2x 2650v2's w/128gb ram in my server currently and would love this as an option! just wondering if i would need an upgrade to get similar performance. currently im giving each gaming machine 8 threads and using antique amd graphics on both and they struggle to say the least. As always enjoying the content watching closely and learning a ton! thanks for all you do!
It doesn't really matter. They will only use the amount of cpu they are using at the time. So it wouldn't be too bad to give each of the machines half of available cores. But when it comes to memory, you need to be a little picky as enabling pci passthrough will disable memory balooning.
fooking life saver mate
pretty neat jow u gpot it all working
I'd like to see you put a Tesla GPU into a system that already has a 5000 or 6000 series Radeon card and have both work. I have gotten errors when using my new 5500 with an older Nvidia 8600 because of Windows 10 getting confused.
This is a good idea for those mining GPUs without any outputs
At least nVidia mining GPUs have NVENC disabled in their firmware, making streaming difficult. AMD mining cards would be great contenders for this though.
I think you should try cooling the Tesla cards, at least the single gpu cards with one of the Arctic GPU coolers or an Kraken G12 AIO setup.
He has in a previous video slapped a giant blower fan on one side
Hi! What You thinks about possibility next gaming scenario: Tesla P6 in virtual machine in slave MXM port and Quadro m4000m as master on laptop P870DM-G (like Optimus mode)? If so, is it possible to run Oculus Rift S?
I have a Tesla M40 that I couldn't figure out how to get to work so anything on getting it functional would be great
I came here to learn about GPU passthrough but I'm just staring at the cat.
Hi man, is there any way to disable the VGX display w/o relying on display settings (they won't save), or activating a quadro license for making the headless display better?
You mentioned that it only works with windows vms. Can you go into why it wont work with linux vms? Also, now that proton is getting more mainstream support, can we use k80s directly on linux without a vm?
Is this supported only on server hardware?
the "first" gpu keeps reporting in correct power usage. (900+ Watt Board power draw).
the "second" gpu works just fine.
its defiantly not drawing anywhere near that. i have a single 450watt PSU just for the k80 (due to the lack of a 8pin eps) and my PSU is not blowing up
I'm trying to do this with a 7700k + ASUS z270f (supports above 4g decoding).
I had to use pcie_acs_override=downstream to split up the IOMMU lanes. without that both GPUs work individually.
Gosh! I shouldve done this before!
For everyone who wanna use this setup to run large language models, use type host in the cpu settings. Otherwise it can not use the shared gpu memory...
Setting cpu to host also makes the vm run faster since you aren't emulating a specific cpu model
as someone looking to upgrade the kiddo setups from ancient Dell boxes, could you recommend a decent ebay server box to get started with one of these? appreciate all the content and guides! this is becoming my favorite channel after LTT :-D
After LTT? Pft.... 😉
Dell R720s include PCIe power most of the time, and make great 2U servers with airflow to support these cards.
@@CraftComputing He meant favorite after BmB
@@CraftComputing thanks for the rec! i'll have to tweet you the current setup while I gameplan the upgrade with your videos
LTT is a clown child that knows almost nothing, his entire corp is affiliate links and shills. Kids watch LTT
I'm trying this with a Tesla P4 but I can't switch from TCC to WDDM with the Quadro drivers. Is it possible to do this with a P4 without the GRiD license?
What is the longest single sentence, involving computers/software,
you can make using mostly abbreviations ? (contest)
(it should be spoken as a single sentence,
not seperate sentences stringed together.)
(the primary score is for the amount of abbreviations.)
Tried this with a Tesla P4 but I can't switch from TCC to WDDM. Does this only work with older cards?
Amazing video!!! There actualy is a waterblock for the tesla k80 on aliexpress.
Yep, because it shares the same PCB as the Titan Z card :-)
Does anyone know if this will work with a Tesla T4? I haven't had success with his other guide video that wasn't for a VM. Edit: when I change the registry for the card by deleting AdapterType and adding EnableMsHybrid, the GPU has an error in device manger.
Time for a easy guide revisit now with the script and higher optics on paravirtualisation?
I wish the two GPU's on the k80 could be bridged with sli to complete ur virtualization hack.
As I understand it will also work without Parsec with Steam remote play?
if my second GPU is not detected in HOST MACHINE(simply because I didnt installed that GPU specific drivers, if I install it will start working, but I cant install because I'm using other Nvidia drivers) soooo ... will I be able to install drivers in Virtual machine? will the virtual machine be able to detect my GPU....which is actually not detected in HOST ?
I was trying exactly this last week then Craft Computing just jumps in and releases that video. Perfect timing man! :)
However I'd like to ask if you had any issues with the M40? I have one and the Nvidia Drivers under windows are throwing Error 12 on me: "This device cannot find enough free resources that it can use. (Code 12)"
Have you seen this before? I'm running a Dell R730 Server with a M40 GPU.
Code 12 is usually a resource issue with PCIe or memory space. Because of the 24GB of VRAM, they don't play well with some motherboards. They should be OK in the R730 though.
@@CraftComputing Hey Jeff. Thanks for the comment. Have you tried your M40 under proxmox and the same Windows VM ? I tried to change the card to a different slot in the motherboard but still no dice...
This is a great option for people who can´t buy any GPUs at the moment