It's also been supported on Intel iGPUs for a very long time. One of the things that isn't discussed so much wrt the topic of GPU virtualization is that the amount of effort that NVidia puts into their driver in their attempt to keep this type of thing locked down on their consumer cards (even though they are capable) is actually the root cause for why doing things like PCIe passthrough of NVidia GPU is such an absolute pain in the ass. NVidia are the most two-faced assholes that ever walked the planet. More of this, Intel, PLEASE
I was thinking about your project watching this, lol. I tried doing virtual desktops at home with 2 S7150x2s I got for cheap a couple years ago hoping for better luck but couldn't solve the same problems you did and accepted it was over. Now I'm just wondering which of those A770s will work
"life changing" is an understatement, I thought we would see this back in 2010. Absolutely insane it has taken this long to make it happen. I had to jump through so many hoops to make Proxmox do Virtual GPUs on my GTX 1060 6GB. Perhaps this will force Nvidia to rethink their policy on VGPUs in a few more years.
rethink? they'll probably increase subscription fee, IT admins don't have power to spend money, its the management, and they'll only pick what their analysts say, and those analysts don't care about a rounding error in their budget but big names like nvidia cloud native.
@@joelv4495- You may or may not know this.. but GPU passthrough is fine if you only want one graphics card powering one virtual machine. Virtual GPUs are when you split up a single graphics card and assign parts of it to different virtual machines.
I'd love to see Intel handle the odd use cases like AMD handles ECC RAM on consumer Ryzen platforms - Enabled, but not certified. Let people mess with it if they want as long as they know it hasn't been given rigorous testing.
Intel do support ECC on consumer CPUs... I'm using ECC RAM with an i5-13500. They've just chosen to artifically limit it to workstation (W680) and server motherboard chipsets. This is purely to make more money... The memory controller is on the CPU so it really shouldn't matter which motherboard chipset you're using.
@@Daniel15au Just to be clear, first of all, it sounds like we agree, and that's why I 'liked' your response. Second of all, I said "platforms," not "CPUs."
this is 100% one of my lofty dreams for my homelab i've been working towards for years but has always been next to impossible or working with jank workarounds. One day i'll have my desktops and gaming centralized in my rack
I purchased a mini-PC with an N100 CPU to run a Jellyfin server in Proxmox. I spent hours and hours trying to get the iGPU pass through working. I was about to give up & try Docker instead. I’m so grateful that this video was made & setup guides were linked in the description. I’ll work through them later, but I can tell at a glance that there were steps I missed because they were not in Jellyfin or Proxmox documentation.
I dream of the day I can dynamically partition my system and GPU and boot linux/windows simultaneously with 1% performance hit. No hal, no paravirt. VMs talk to the hardware. We don't have to put up with latency in 202X
This is kind of what I want for a whole house gaming setup. I want a big PC rack mounted in my comm closet and then every TV/monitor is just a thin station. Even if I couldn't slice one GPU up but had to dedicate a GPU to each VD. I haven't figured out how to make a setup that's acceptable for gaming. Local is still always better by a long shot.
@@lyserberg You are aware that a vast majority of games don't use anti cheat right? Your entire point falls flat on it's face if someone just doesn't play zoomer shooters lol
@@Real-Name..MaqavoySo this is the setup I have for my main PC. The only Anti-cheat game that I am aware of that I have used is the master chief collection but it worked fine. Steam hasn't given me any issues with the setup. Currently I run an old epyc 7532 on truenas scale and am passing a 4070 through to a windows 10 VM. It works fine. Don't try with an AMD GPU though because the drivers fail no matter how I have tried to set it up.
Stream to 5 separate gpu accelerated clients around my house and support AV1 encode/decode for media server support? This is like a dream case for the tech dad/mom house build. Also, possibly a very powerful streamer house setup...
@@wayland7150 It's hard to find a consumer platform with 2 x16 slots. The best you can hope for is 2 slots in x8 mode and you won't have any slots left for I/O. 1 slot powering 4 VMs is much more useful.
They will fight this to the teeth. The moment intel or nvidia enables this on a consumer grade a big chunk of the market will save money on this…. We can dream tho
Our server admins in the other department will never admit this to be true... Because Linux hypervisor environments are "cobbled together". And their great Citrix VDI Test machine based on VMware and ancient Nvidia Quadro GPUs was a grandiose failure because... You guessed it... Subscription fees!!!
This is what everyone should have. Basically a home server that you could split up between multiple users in a home. In that way each endpoint only needs to be as powerful as a cellphone. Personally I am using truenas scale right now but I hope they can flush out their virtualization capabilities but I am really excited that proxmox is getting these capabilities.
I would love to have this to separate my gaming and machine learning environments, but any sort of virtualization vs anti-cheat is a no-go. I don't want to do some weird Frankenstein's monster of Windows as host and hope that GPU passthrough to Linux works, and I don't want to deal with WSL because eventually I'll run into some issue.
In order for this to become popular, hardware manufacturers need to start making guarantees to virtual machines about whether they can access guest memory, and then you have to get the developers on board and to trust it, as well as the OS vendors, etc.
@@supdawg7811I need to try running some more anti-cheat games on my system. Currently I have only tried the master chief collection but I don't really like playing on the internet with random people. Just not the kind of games I like. Currently my main "desktop" though is a windows 10 system I have passed a 4070 super through to and it works pretty well.
@@supdawg7811Yeah I do think that the software ecosystem needs to evolve a bit because currently getting a performant virtualized systems up an running can be a bit of a pain. Previously I was running a system like you said with a windows server based system but when I tried upgrading my GPU from a 1060 to a 6800 AMDs installer told me that they didn't like that I was using their GPU on windows server. I used windows server because it was also serving as my file server and I got the license for free from school. I then decided to Install truenas scale, since the primary purpose here was a file server, and pass the 6800 to the VM but I still couldn't get the AMD drivers to install properly. I tried switching to Proxmox which allowed me to try with windows 10 and 11 but had the same issue on both. Tested passing my old 1060 to a VM and found that it worked with Nvidia GPUs so switched back to truenas scale, bought a new 4070 super, and have been using that for a couple of months now. I love it but the software restrictions AMD had in place have really pissed me off.
Its been so many years! Please Intel release a consumer card with enabled SR-IOV. You don’t need to support it. Just put a warning in the driver (like overclocking on motherboards) and let us tinker to our heart’s content!!
Been looking into VDI for a few years now to potentially cut power costs, more specifically the number of running systems. Really glad this has got some attention as it may make it a more viable system letting companies know there are a casual use cases for these cards.
Want! After spending a weekend of driver and licencing hell getting vGPU up on a Tesla P4 I would happily pay to never have to deal with licencing servers and the ability to use drivers from my distros package repository.
I am looking to buy a used card since I am working on getting my certifications for OpenShift/VMware/Citrix given that I am or have worked and maybe be deployed again at clients that use it. This is awesome for learning at home and getting actual hands-on use and learning with this technology. I never bothered with buying a used NVIDIA or AMD card for licensing reasons or in the case of AMD, general support on Citrix or VMware Horizon for example. Anyway, many thanks for this video Wendell, you gave us a sneak peak of this on the L1 Linux channel but to see this is full form is just amazing!
We're buying at least $10M in Dell this year without alot of that being 2u VSAN ready nodes for Vmware with GPU slots standard. Haven't purchased many GPUs but the few I've seen have been Nvidia a40 (I think?) data center GPUs. We have some old VDI on ancient Vblocks so may need an upgrade here soon. Would be interested in what my Dell rep has or our VAR.
I've got a buddy at another VAR says they have the Flex 140 in their catalogue, on backorder and not shipping but eventually available for about $2400.
Eyes open for one for my home lab VMware Horizon implementation. Already have a Tesla P4 passed through to a Plex server, time to give my virtual desktops some love.
Man the Gandalf thing is a proper throwback, I remember that meme from the early days haha Nice to see these intel gpu capabilities though, interesting, I'll be keeping an eye on that next time we go through our refresh, see if I can't procure something like this for testing in the meantime
It is genuinely amazing that the VDI industry has been so ridiculous that $400/seat for just the virtual part (no thin client, no peripherals) is a steal.
Simple homelab setup here with no accelerated stuff at all. No plans to upgrade just yet but if this works with a standard ARC card or the flex cards are reasonably priced then wow I may have to look into adding this in to my current setup.
My lord, I'm actually exited about getting back into server hardware again. This would make something like an active display wall possible where you could interact with it on 100's of displays at an event.
It will make a big difference down the road when people upgrade the districts computer systems. Pushing 30 desktops 3D is amazing. I am hoping for a arc will enable this feature in the future on desktops.
I'm using SR-IOV on my Unraid home server with i5-14500 (UHD 770 iGPU) - it is a plugin (in beta) and it works great so far, I just set how many "VFs" I want (0-7) and I can pass it to any VM. I didn't even need to do anything in CLI, just downloaded a plugin from CA apps - imagine that, using a GUI for things on Linux, half of the Truenas forum would be so mad.
I enabled sr-iov on a low power Intel N100 mini pc with proxmox. Works a charm for encoding jellyfin streams (on lxc) and windows 11 remote desktops (qemu) simultaneously
Kudos to Intel on this. I'm solidly red team for compute and red or green team for GPU, but I appreciate anything thing that is performant and cost effective.
Thank you for continued great content. Just a bad prosumer here….but I have made a career in IT just based on videos from the likes of you and Linus, etc. high five 🖐️ from Kansas City!
Every Intel project I've invested in or bought components for in the past 3+ years was canceled by Intel. From FPGA's to RISCV platforms. I laughed so hard last week when I found out that Altera is now a standalone company again, really Intel? Intel is like the koolaid man that smashes through your walls and shouts "Guess Who's Back!"
@@edc1569 At least they had excuse that there was no killer app back then. I wonder how much the industry shrugs when Intel launches a new product because everyone knows that within a couple years they will take a hammer to it. You can't even escape this in shipping products. I had an order for FPGA's that I'd ordered from mouser. 3 months later I got an email saying all my orders were canceled as Intel is shipping the darn things to the DOD. They weren't even the high end parts.
I just started doing this on proxmox a week ago with a raptor lake igpu and it works fantastic. My docker LXC has video acceleration and so does my blue iris windows VM. Honestly if intel enabled this for the a770 or even the upcoming battlemage gpus I would definitely consider switching from my 3080 as I'd love to have linux as my primary gaming OS on my desktop with a window VM just in case it's needed. Single gpu IOMMU passthrough is too janky for my taste, sr-iov would solve basically every complaint I have.
Can you test that SR-IOV alongside enabling nested virt on that VM in order tu deploy WSA/WSL2 environments? Because when i do, i have to change CPU passthrough model and then SRIOV driver gives me error 43. Has to be exact model, but since nested virt doesn't work with, i'm in a fire circle. Thanks.
@@Pacho18 Sorry no. I currently don't have any use case that needs that and I don't currently have the time in the next few weeks to test that out for giggles.
Maybe you can answer something for me. I have a I7-11700 as well as an Nvdia RTX3070TI in a box running Blue Iris and CodeProject AI in a W11 VM with GPU pass thru of the 3070TI. I have Home Assistant running in another VM. Prior to this, I was just running without Home Assistant on bare metal W11. With Bare metal, the iGPU in the 11700 was doing most of the AI work as well as providing Hardware Encoding of the camera streams, roughly 10 cameras and I was able to do this on about 70-90 watts. After converting the system to Proxmox, the power draw is more like 140-160 watts, due to the RTX3070TI handling all acceleration requirements for the VM. My question to you is: Would my power requirements go down to switch over to the iGPU and virtualize it, and not use the RTX3070TI ? I know I could try this out, but it’s an awful lot of work to get everything switched over only to find out I haven’t really saved any power. Normally I don’t really care about the power requirements except for UPS planning in the event of power loss, then the ~50 watt difference gets a lot more significant. Thanks in advance! JC
I would love to see a video demonstrating alder lakes igpu SR-IOV capabilities and setup process, maybe a cubesOS like functionality with hardware accelerated vga on the cheap would be possible with this?
This is huge. I would pick one up, but 5k is a bit much for just homelab use. I hope it gets more popular in the enterprise space, so in a few years the price will be suitable for enthusiasts.
I am prosumer and I have been looking for a flex 170 on eba recently, but it seems to be to new and unpopular for good deals to appear on ebay.... at least within the EU without crazy import fees
I'm really looking forward for the video about SR-IOV with desktop GPUs! Unfortunately Intel Flex is quite out of reach for the projects im working on 😅 There is also the VirGL GPU type in Proxmox which, as I understand it, enables sharing of GPU resources without explicit GPU support. I never got it working with a significant performance increase. Can someone tell me if it even works?
@@Pacho18 wow thanks 😂 that's something I got to myself. I have a test machine with a GPU and Linux guests that use VirGL, but I'm not that deep into how "rendering" works so that I can confirm that it is really using the GPU. Maybe I should also try it again in PVE 8.1.
Their gaming GPUs are also the best for price to performance and are much closer to NVIDIA features than AMD. Their new fabs are monsters too. Intel is about to go big again, i think.
Intel have been fairly good about software support for many years, the issue has always been gating hardware, and that's still very much their MO when you look at how much a Flex 170 costs compared to a more or less identical a770 from a raw silicon perspective
Now thinking much more about the future, with this technology we can access high processing power from anywhere, and going further, we can have our own open source AI.
But does it do AV1? 😅 Even if it doesn't... If I understand correctly it allows to encode 32*1080p streams or 8*4K streams? The ability to stream to basically all the platforms you'd want to stream (YT, Twitch, FB, TikTok, Vimeo, Twitter, Instagram, DLive, Trovo, Kick, LinkedIn 🤣) AT THE SAME TIME and with DIFFERENT SETTINGS EACH is wild.
@level1techs You mentioned a per seat cost of $400. Where did this number come from? Are you saying that the single server, with 16 clients or charging the customer $400/per VDI.
The only thing that I'd like more than subscription-free vGPUs is ones that are more like CPU virtualization (in feature set) because it's pretty nice that you can dynamically limit how much power a VM has access to or just let them all have full access but limit your usage inside them. Having to slice a GPU down into tiny chunks is unfortunate if you suddenly want to shift power to one VM and away from another. EDIT: Since it was mentioned, I mean supported on a hardware level (if such a thing is possible), so that they don't need to do software GPU partitioning which from what I understand lowers performance by quite a bit.
Our CAD users are looking for a solution for using Revit and Civil 3D remotely. There are DISA reasons that make it difficult but I'll add this to the list.
This is cool. But it doesn't solve the issue that the vast majority of computers won't be able to do this due to lack of SR-IOV hardware support. However we could in theory solve this in software without needing SR-IOV, by using paravirtualization with vulkan API. Then we can have GPU accelerated VMs on Vulkan capable GPUs. I think it's already possible to do this for linux guests with virgl venus protocol. Maybe it would be possible to use the venus protocol to develop a windows guest driver as well, and using DXVK, vkd3d and Zink to get all the APIs we need. I think such a solution could perform pretty well.
Ok I know this is stupid of me to comment multiple times on the same video but I love this topic. I wonder if we could lower the latency of VDI if we were to stop encoding and decoding all of the video. I know we do this because it isn't practical to send uncompressed video streams over a network. 48Gbps that HDMI 2.1 provides to hit 4k120fps isn't doable on modern networks. Or is it? The thing is that you, yes you, can buy a 100Gbps switch from ebay for about $1000. Theoretically that is plenty of bandwith to allow an uncompressed video stream as well as allow bandwidth for any peripherals you may have. Even Wifi 7 has a theoretical maximum of 46Gbps and should allow for a direct connection wirelessly. If we were to upgrade the networking in our homes to fiber could we achieve the holy grail of VDI by having low latency access to virtual systems from anywhere in our home? I would love to see us work towards that.
Personally I've been wanting to do a hobbyist cloud gaming setup for years, most importantly with old lan games to play with my friends. My dream is just to have them remote in from any device (mainly laptops that connect to my LAN) and play all those old games together. This way they don't have to download anything, configure anything, they don't have to fiddle with incompatibilities (win10 sucks for old games), and the VDIs they jump into already have the workarounds and the time investment on getting everything working. So far I only experienced failure, and I swore not to go back to any of that nvidia grid crap, which is nothing else but pain to set up, and never ends up working (for a hobbyst setup from a dude like me that will never pay the license).
Not to belittle this, I am excited at the future prospects, but at $2000 a card, it's more financially viable to buy a PCI-E bifurcation card and multiple RTX cards with PCI-E passthrough instead. Wouldn't be as compact if you were looking to colocate however.
@@Pacho18The PCIe switch card is effectively its own PCIe endpoint, so if you can get a gen 5 device it will interact with the host system at true gen 5 speeds even if the downstream devices are gen4. Very difficult to get in the consumer space but in theory this is the only way to get the full bandwidth of a gen 5 slot if using gen 4 endpoint devices (eg a 16 lane to 32 lane gen 5 switch will happily provide simultaneous full speed access to 8 gen 4 NVMe drives because the gen 4 drives only force the connection between the switch and the drive down to gen 4, the upstream link is separate and can run at gen 5 as long as the host can).
AMD also has this capability for ESXi, Citrix, and Linux (this is public information - MxGPU - and has existed for a long time). But the up to date tooling to enable it isn't outright available. 😔
It should be titled "Subscription-free GPU accelerated VDI", because we already have subscription free VDI
Good call! ~Editor Autumn
@@Level1Techs I swore it was "Intel introduces Subscription GPU"
It's also been supported on Intel iGPUs for a very long time.
One of the things that isn't discussed so much wrt the topic of GPU virtualization is that the amount of effort that NVidia puts into their driver in their attempt to keep this type of thing locked down on their consumer cards (even though they are capable) is actually the root cause for why doing things like PCIe passthrough of NVidia GPU is such an absolute pain in the ass. NVidia are the most two-faced assholes that ever walked the planet.
More of this, Intel, PLEASE
"for now" tm
Hey what are some free VD I can use?
YAAAAAY!!!!!
time to make another episode in the long lived "we have vdi at home" series, i guess
I was thinking about your project watching this, lol. I tried doing virtual desktops at home with 2 S7150x2s I got for cheap a couple years ago hoping for better luck but couldn't solve the same problems you did and accepted it was over. Now I'm just wondering which of those A770s will work
a770 gpu gaming incoming?
Bonus points for epic sax guy. -the guy who worked with you on Nvidia GRID some 5+ years back
"life changing" is an understatement, I thought we would see this back in 2010. Absolutely insane it has taken this long to make it happen. I had to jump through so many hoops to make Proxmox do Virtual GPUs on my GTX 1060 6GB. Perhaps this will force Nvidia to rethink their policy on VGPUs in a few more years.
rethink? they'll probably increase subscription fee, IT admins don't have power to spend money, its the management, and they'll only pick what their analysts say, and those analysts don't care about a rounding error in their budget but big names like nvidia cloud native.
Intel is for more well known.
TrueNAS Scale has GPU passthrough to VMs in the GUI! I'm a recent convert from Proxmox, and I'm gonna start playing with GPU passthrough soon.
@@joelv4495- You may or may not know this.. but GPU passthrough is fine if you only want one graphics card powering one virtual machine. Virtual GPUs are when you split up a single graphics card and assign parts of it to different virtual machines.
@@hycron1234 oh gotcha. Makes sense that a totally different process is required.
I can't believe the title reads "subscription free" and we're all excited about it. DARKEST TIMELINE CONFIRMED.
I'd love to see Intel handle the odd use cases like AMD handles ECC RAM on consumer Ryzen platforms - Enabled, but not certified. Let people mess with it if they want as long as they know it hasn't been given rigorous testing.
Intel do support ECC on consumer CPUs... I'm using ECC RAM with an i5-13500. They've just chosen to artifically limit it to workstation (W680) and server motherboard chipsets. This is purely to make more money... The memory controller is on the CPU so it really shouldn't matter which motherboard chipset you're using.
@@Daniel15au Just to be clear, first of all, it sounds like we agree, and that's why I 'liked' your response. Second of all, I said "platforms," not "CPUs."
@@eldibs sorry, I misread "platforms" as "CPUs" in your comment!
@@Daniel15au My good bromigo, it's okay, an honest mistake. Also, your extra info was interesting.
I believe ecc is deliberately disabled on Intel cpu's, so they can ask for a premium to have it enabled. AMD leaves it enabled
this is 100% one of my lofty dreams for my homelab i've been working towards for years but has always been next to impossible or working with jank workarounds. One day i'll have my desktops and gaming centralized in my rack
Wendell is that you? You're the whole reason my business took off, you taught me so much
I purchased a mini-PC with an N100 CPU to run a Jellyfin server in Proxmox.
I spent hours and hours trying to get the iGPU pass through working. I was about to give up & try Docker instead.
I’m so grateful that this video was made & setup guides were linked in the description.
I’ll work through them later, but I can tell at a glance that there were steps I missed because they were not in Jellyfin or Proxmox documentation.
Did you end up getting SR-IOV working with the N100 iGPU?
Wow this is a real game changer we will get one for our internal proxmox server. Thanks for this information. 🙏🏿🙏🏿🙏🏿
The moment I can buy a consumer GPU with SR-IOV, I'm in.
Yes please. I have a spare slot just waiting in my server...
I dream of the day I can dynamically partition my system and GPU and boot linux/windows simultaneously with 1% performance hit. No hal, no paravirt. VMs talk to the hardware. We don't have to put up with latency in 202X
So you're buying an A770?
@@hugevibez Pretty sure that's not officially supported at this point. Intel and Nvidia both teased the option, but that obviously didn't pan out. :-/
They know that, and want you spending enterprise monies instead :/
Man Wendell i can't tell you how thankful i am for your videos, your info, and everyone at the L1 forums! A million thank yous!!
This is dynamite! Thanks for this Wendell!
This is kind of what I want for a whole house gaming setup. I want a big PC rack mounted in my comm closet and then every TV/monitor is just a thin station. Even if I couldn't slice one GPU up but had to dedicate a GPU to each VD. I haven't figured out how to make a setup that's acceptable for gaming. Local is still always better by a long shot.
You could never do it for gaming anyway. You'd fail every bit integrity check straight out the door by even the lousiest of anti-cheat systems.
@@lyserbergthat was my question is how do you plan to circumvent anti virtualization measures taken by anti cheats
@@lyserberg You are aware that a vast majority of games don't use anti cheat right? Your entire point falls flat on it's face if someone just doesn't play zoomer shooters lol
i never had an issue playing games in visualised environment! even ubisoft games work (there just funny about getting too many cpu!) @@lyserberg
@@Real-Name..MaqavoySo this is the setup I have for my main PC. The only Anti-cheat game that I am aware of that I have used is the master chief collection but it worked fine. Steam hasn't given me any issues with the setup. Currently I run an old epyc 7532 on truenas scale and am passing a 4070 through to a windows 10 VM. It works fine. Don't try with an AMD GPU though because the drivers fail no matter how I have tried to set it up.
Get this to consumers! Maybe not 32 vdi but allow to have like 5, and AV1 encoding for cheap and it will be amazing
Stream to 5 separate gpu accelerated clients around my house and support AV1 encode/decode for media server support? This is like a dream case for the tech dad/mom house build.
Also, possibly a very powerful streamer house setup...
Alternatively fit several discreet GPUs and pass them through.
@@wayland7150 It's hard to find a consumer platform with 2 x16 slots. The best you can hope for is 2 slots in x8 mode and you won't have any slots left for I/O. 1 slot powering 4 VMs is much more useful.
They will fight this to the teeth. The moment intel or nvidia enables this on a consumer grade a big chunk of the market will save money on this…. We can dream tho
As a Horizon admin that had to give up his K1 and K2 Grid cards this warms my cold heart.
god damn it MAKE THE ARC VIDEO!!! i've watched this video SO MANY TIMES!! I CANNNOT WAITTTTTTT!!!
Our server admins in the other department will never admit this to be true... Because Linux hypervisor environments are "cobbled together".
And their great Citrix VDI Test machine based on VMware and ancient Nvidia Quadro GPUs was a grandiose failure because... You guessed it... Subscription fees!!!
This is what everyone should have. Basically a home server that you could split up between multiple users in a home. In that way each endpoint only needs to be as powerful as a cellphone. Personally I am using truenas scale right now but I hope they can flush out their virtualization capabilities but I am really excited that proxmox is getting these capabilities.
I would love to have this to separate my gaming and machine learning environments, but any sort of virtualization vs anti-cheat is a no-go. I don't want to do some weird Frankenstein's monster of Windows as host and hope that GPU passthrough to Linux works, and I don't want to deal with WSL because eventually I'll run into some issue.
In order for this to become popular, hardware manufacturers need to start making guarantees to virtual machines about whether they can access guest memory, and then you have to get the developers on board and to trust it, as well as the OS vendors, etc.
You're basically looking at extensive use of TPM, maybe per-guest encrypted memory, etc.
@@supdawg7811I need to try running some more anti-cheat games on my system. Currently I have only tried the master chief collection but I don't really like playing on the internet with random people. Just not the kind of games I like. Currently my main "desktop" though is a windows 10 system I have passed a 4070 super through to and it works pretty well.
@@supdawg7811Yeah I do think that the software ecosystem needs to evolve a bit because currently getting a performant virtualized systems up an running can be a bit of a pain. Previously I was running a system like you said with a windows server based system but when I tried upgrading my GPU from a 1060 to a 6800 AMDs installer told me that they didn't like that I was using their GPU on windows server. I used windows server because it was also serving as my file server and I got the license for free from school. I then decided to Install truenas scale, since the primary purpose here was a file server, and pass the 6800 to the VM but I still couldn't get the AMD drivers to install properly. I tried switching to Proxmox which allowed me to try with windows 10 and 11 but had the same issue on both. Tested passing my old 1060 to a VM and found that it worked with Nvidia GPUs so switched back to truenas scale, bought a new 4070 super, and have been using that for a couple of months now. I love it but the software restrictions AMD had in place have really pissed me off.
Wendel is a god of nerds, such a beast
Its been so many years! Please Intel release a consumer card with enabled SR-IOV. You don’t need to support it. Just put a warning in the driver (like overclocking on motherboards) and let us tinker to our heart’s content!!
Fantastic video! Proxmox for the win!
Been looking into VDI for a few years now to potentially cut power costs, more specifically the number of running systems. Really glad this has got some attention as it may make it a more viable system letting companies know there are a casual use cases for these cards.
Ha! We are literally looking at deploying a new VDI solution. Glad I remembered you uploaded this video.
So what your saying is I should buy the "correct" a770 for my proxmox cluster....very well I shall visit the forums for the "one".😅
I've never been able to find the actual files to flash for it.
@@ionstorm66 well damn lol
I wonder if this could work on cheaper cards like the a380 as well?
@@TomMorris1 I happen to have a a380 in one of my nodes now. 🤔
@@TomMorris1I think the 380 is nearly as strong for this type of task?
Want!
After spending a weekend of driver and licencing hell getting vGPU up on a Tesla P4 I would happily pay to never have to deal with licencing servers and the ability to use drivers from my distros package repository.
Your channel is awesome. I have been looking for a solution like this for years.
Absolutely loved the new intro. Excellent video, the future of vdi is looking accelerated!
I am looking to buy a used card since I am working on getting my certifications for OpenShift/VMware/Citrix given that I am or have worked and maybe be deployed again at clients that use it. This is awesome for learning at home and getting actual hands-on use and learning with this technology. I never bothered with buying a used NVIDIA or AMD card for licensing reasons or in the case of AMD, general support on Citrix or VMware Horizon for example.
Anyway, many thanks for this video Wendell, you gave us a sneak peak of this on the L1 Linux channel but to see this is full form is just amazing!
This is great stuff especially regarding Proxmox! Excelling content. Thanks for making it happen!
I read and reread the title like 3-4 times and read “subscription fee” every time. ADHD for the win.
It needs a dash. I'm a heavy dash-user 😄
ADHD? What?
@@skunkwerx9674 attention deficit with hyperactivity disorder. I don't remember the actual acronym
Oh my god this is GOOD news! Thanks for unraveling this awesome development Wendell
the best part of the video was the datacenter ASMR... it brings back so many memories...
We're buying at least $10M in Dell this year without alot of that being 2u VSAN ready nodes for Vmware with GPU slots standard. Haven't purchased many GPUs but the few I've seen have been Nvidia a40 (I think?) data center GPUs. We have some old VDI on ancient Vblocks so may need an upgrade here soon. Would be interested in what my Dell rep has or our VAR.
I've got a buddy at another VAR says they have the Flex 140 in their catalogue, on backorder and not shipping but eventually available for about $2400.
Please keep us updated on this, extremely interested in testing.
Oh thank god I am not the only one nerding out over this.
I just bought an Intel ARC A60 love it!
Eyes open for one for my home lab VMware Horizon implementation. Already have a Tesla P4 passed through to a Plex server, time to give my virtual desktops some love.
I've been waiting for video like this
Man the Gandalf thing is a proper throwback, I remember that meme from the early days haha
Nice to see these intel gpu capabilities though, interesting, I'll be keeping an eye on that next time we go through our refresh, see if I can't procure something like this for testing in the meantime
LOL. Thanks for that intro. Gunna have that Eurovision song in my head for another week. HAHAHA!
Impressive, people like to hate on intel but I'm really excited about their GPUs for workflows in video and 3D. Go team blue! 🔵
It is genuinely amazing that the VDI industry has been so ridiculous that $400/seat for just the virtual part (no thin client, no peripherals) is a steal.
"... we didn't account for this use case..."
that brought back some tears in my eyes...
Simple homelab setup here with no accelerated stuff at all. No plans to upgrade just yet but if this works with a standard ARC card or the flex cards are reasonably priced then wow I may have to look into adding this in to my current setup.
flex cards are not reasonably priced for consumers.
My lord, I'm actually exited about getting back into server hardware again. This would make something like an active display wall possible where you could interact with it on 100's of displays at an event.
It will make a big difference down the road when people upgrade the districts computer systems. Pushing 30 desktops 3D is amazing. I am hoping for a arc will enable this feature in the future on desktops.
I'm using SR-IOV on my Unraid home server with i5-14500 (UHD 770 iGPU) - it is a plugin (in beta) and it works great so far, I just set how many "VFs" I want (0-7) and I can pass it to any VM. I didn't even need to do anything in CLI, just downloaded a plugin from CA apps - imagine that, using a GUI for things on Linux, half of the Truenas forum would be so mad.
Which plug in?
Yes, any more info on this you could provide would be helpful.
Dude half the comment sections was asking "770 sr-iov when". You gotta spill the sauce
If you follow the how to on the forum it works same for minisforum and uhd 770 as shown here. Video on l1 linux soon
@@pauljones9150 not Arc 770, but UHD 770 - integrated GPU on i5-14500. In CA apps it is called "Intel i915 SR-IOV" from giganode.
Love the jump scare, reminds me of that old classroom vid
we're doing testing with the flex gpus on VxRail currently in the 16G Dells.
I really hope that the market is hot again with 3 competitors instead of just one company monopoly market
I enabled sr-iov on a low power Intel N100 mini pc with proxmox. Works a charm for encoding jellyfin streams (on lxc) and windows 11 remote desktops (qemu) simultaneously
Kudos to Intel on this. I'm solidly red team for compute and red or green team for GPU, but I appreciate anything thing that is performant and cost effective.
The moment Wendell, is Geeked all the way out! They finally caught-up, Wendell.
Thank you for continued great content. Just a bad prosumer here….but I have made a career in IT just based on videos from the likes of you and Linus, etc. high five 🖐️ from Kansas City!
I'm intrigued by the accidental 30 virtual functions on A770. Any news on that?
Every Intel project I've invested in or bought components for in the past 3+ years was canceled by Intel. From FPGA's to RISCV platforms. I laughed so hard last week when I found out that Altera is now a standalone company again, really Intel? Intel is like the koolaid man that smashes through your walls and shouts "Guess Who's Back!"
Remember that really innovative camera thing they built and then scrapped while everyone in industry was getting their heads around it, realsense.
@@edc1569 At least they had excuse that there was no killer app back then. I wonder how much the industry shrugs when Intel launches a new product because everyone knows that within a couple years they will take a hammer to it. You can't even escape this in shipping products. I had an order for FPGA's that I'd ordered from mouser. 3 months later I got an email saying all my orders were canceled as Intel is shipping the darn things to the DOD. They weren't even the high end parts.
@@edc1569 My response was censored by the Google/RUclips overlords. I despair this place.
@@edc1569lidar?
I just started doing this on proxmox a week ago with a raptor lake igpu and it works fantastic. My docker LXC has video acceleration and so does my blue iris windows VM.
Honestly if intel enabled this for the a770 or even the upcoming battlemage gpus I would definitely consider switching from my 3080 as I'd love to have linux as my primary gaming OS on my desktop with a window VM just in case it's needed. Single gpu IOMMU passthrough is too janky for my taste, sr-iov would solve basically every complaint I have.
Can you test that SR-IOV alongside enabling nested virt on that VM in order tu deploy WSA/WSL2 environments? Because when i do, i have to change CPU passthrough model and then SRIOV driver gives me error 43. Has to be exact model, but since nested virt doesn't work with, i'm in a fire circle.
Thanks.
@@Pacho18 Sorry no. I currently don't have any use case that needs that and I don't currently have the time in the next few weeks to test that out for giggles.
Maybe you can answer something for me. I have a I7-11700 as well as an Nvdia RTX3070TI in a box running Blue Iris and CodeProject AI in a W11 VM with GPU pass thru of the 3070TI. I have Home Assistant running in another VM. Prior to this, I was just running without Home Assistant on bare metal W11. With Bare metal, the iGPU in the 11700 was doing most of the AI work as well as providing Hardware Encoding of the camera streams, roughly 10 cameras and I was able to do this on about 70-90 watts. After converting the system to Proxmox, the power draw is more like 140-160 watts, due to the RTX3070TI handling all acceleration requirements for the VM. My question to you is: Would my power requirements go down to switch over to the iGPU and virtualize it, and not use the RTX3070TI ? I know I could try this out, but it’s an awful lot of work to get everything switched over only to find out I haven’t really saved any power. Normally I don’t really care about the power requirements except for UPS planning in the event of power loss, then the ~50 watt difference gets a lot more significant. Thanks in advance! JC
The Flex 170 is a bit too expensive for my hobbyist needs, but I would love an A770 I could split into 4 (8 would be amazing).
Insanely cool 👍
Dream come true
I would love to see a video demonstrating alder lakes igpu SR-IOV capabilities and setup process, maybe a cubesOS like functionality with hardware accelerated vga on the cheap would be possible with this?
Yes but how do we get the Flex 170s? I went to the Supermicro site but it just shows the server and config but no option for Flex 170s? What gives?
Just one of these would solve half my virtual machine needs…pretty nice. Hopefully we get to see it in more affordable (for the consumer) someone.
i like the subtle soundtrack
This is exciting!
This is huge. I would pick one up, but 5k is a bit much for just homelab use. I hope it gets more popular in the enterprise space, so in a few years the price will be suitable for enthusiasts.
especially as it is a Arc A770 (what £300!!!! and no cooling fan!)
Thanks Wendell!
8:02 wow what a shot! 🤣
HYPE HYPE HYPE HYPE (IM EXCITED)
I am prosumer and I have been looking for a flex 170 on eba recently, but it seems to be to new and unpopular for good deals to appear on ebay.... at least within the EU without crazy import fees
Sounds good. Now I want a Minisforum MS-01 for the iGPU SRIOV use in a Jellyfin VM.
If there was a hall of IT Wendel should be in it!
Had to check my phone post intro, as the epic saxman is my phones ringtone XD
I'm really looking forward for the video about SR-IOV with desktop GPUs! Unfortunately Intel Flex is quite out of reach for the projects im working on 😅 There is also the VirGL GPU type in Proxmox which, as I understand it, enables sharing of GPU resources without explicit GPU support. I never got it working with a significant performance increase. Can someone tell me if it even works?
VirGL only works with linux guests.
@@Pacho18 wow thanks 😂 that's something I got to myself. I have a test machine with a GPU and Linux guests that use VirGL, but I'm not that deep into how "rendering" works so that I can confirm that it is really using the GPU. Maybe I should also try it again in PVE 8.1.
wow & thank you
it'd be really neat if this could ever get back supported to consumer amd cards. could see some nerdy home networks that would love this
Intel coming to save the day. Absurdity at it's finest. We live in weird times.
Their gaming GPUs are also the best for price to performance and are much closer to NVIDIA features than AMD.
Their new fabs are monsters too.
Intel is about to go big again, i think.
Intel have been fairly good about software support for many years, the issue has always been gating hardware, and that's still very much their MO when you look at how much a Flex 170 costs compared to a more or less identical a770 from a raw silicon perspective
Now thinking much more about the future, with this technology we can access high processing power from anywhere, and going further, we can have our own open source AI.
But does it do AV1? 😅
Even if it doesn't... If I understand correctly it allows to encode 32*1080p streams or 8*4K streams? The ability to stream to basically all the platforms you'd want to stream (YT, Twitch, FB, TikTok, Vimeo, Twitter, Instagram, DLive, Trovo, Kick, LinkedIn 🤣) AT THE SAME TIME and with DIFFERENT SETTINGS EACH is wild.
It was infact first hardware to do av1..in2022. Just unobtanium
@@Level1Techs yikes
It's crazy how Intels graphics division has become this incredible gpu underdog
@level1techs You mentioned a per seat cost of $400. Where did this number come from? Are you saying that the single server, with 16 clients or charging the customer $400/per VDI.
Probably hardware cost divided by max seat count, these things aren't cheap
This is super cool technology.
The only thing that I'd like more than subscription-free vGPUs is ones that are more like CPU virtualization (in feature set) because it's pretty nice that you can dynamically limit how much power a VM has access to or just let them all have full access but limit your usage inside them. Having to slice a GPU down into tiny chunks is unfortunate if you suddenly want to shift power to one VM and away from another.
EDIT: Since it was mentioned, I mean supported on a hardware level (if such a thing is possible), so that they don't need to do software GPU partitioning which from what I understand lowers performance by quite a bit.
Love the Gandalf reference.
You got Epic Sax Guy, huh? Was the first song my daughter learned to play when she started paying sax.
Its a beautiful thing.
If Citrix qualified this, how will XCP-Ng work?
2:52 oh thank god . The fan noise wasn’t in my head
Wow, you have been wanting this for awhile. Its Intel.
Where can we buy one of those cards? Are there any plans to sell them to us mortals?
Subscription free…………………………for now. I see the strat!
Our CAD users are looking for a solution for using Revit and Civil 3D remotely. There are DISA reasons that make it difficult but I'll add this to the list.
This is cool. But it doesn't solve the issue that the vast majority of computers won't be able to do this due to lack of SR-IOV hardware support. However we could in theory solve this in software without needing SR-IOV, by using paravirtualization with vulkan API. Then we can have GPU accelerated VMs on Vulkan capable GPUs. I think it's already possible to do this for linux guests with virgl venus protocol. Maybe it would be possible to use the venus protocol to develop a windows guest driver as well, and using DXVK, vkd3d and Zink to get all the APIs we need. I think such a solution could perform pretty well.
Ok I know this is stupid of me to comment multiple times on the same video but I love this topic. I wonder if we could lower the latency of VDI if we were to stop encoding and decoding all of the video. I know we do this because it isn't practical to send uncompressed video streams over a network. 48Gbps that HDMI 2.1 provides to hit 4k120fps isn't doable on modern networks. Or is it? The thing is that you, yes you, can buy a 100Gbps switch from ebay for about $1000. Theoretically that is plenty of bandwith to allow an uncompressed video stream as well as allow bandwidth for any peripherals you may have. Even Wifi 7 has a theoretical maximum of 46Gbps and should allow for a direct connection wirelessly. If we were to upgrade the networking in our homes to fiber could we achieve the holy grail of VDI by having low latency access to virtual systems from anywhere in our home? I would love to see us work towards that.
Personally I've been wanting to do a hobbyist cloud gaming setup for years, most importantly with old lan games to play with my friends. My dream is just to have them remote in from any device (mainly laptops that connect to my LAN) and play all those old games together. This way they don't have to download anything, configure anything, they don't have to fiddle with incompatibilities (win10 sucks for old games), and the VDIs they jump into already have the workarounds and the time investment on getting everything working. So far I only experienced failure, and I swore not to go back to any of that nvidia grid crap, which is nothing else but pain to set up, and never ends up working (for a hobbyst setup from a dude like me that will never pay the license).
1) How much does a Flex 170 cost?
2) where can i buy one?
3) how many plex streams could it run?
1) A lot
2) eBay, or preintegrated in a Supermicro machine (the latter potentially being better value for money overall)
3) 32 iirc
Not to belittle this, I am excited at the future prospects, but at $2000 a card, it's more financially viable to buy a PCI-E bifurcation card and multiple RTX cards with PCI-E passthrough instead. Wouldn't be as compact if you were looking to colocate however.
Electrical consumption would be a lot higher...but I use a solution for the issues that arise from that in my game server.
How does thag PCi-e bifurcation works, having into account different versions of the standard?
@@Pacho18The PCIe switch card is effectively its own PCIe endpoint, so if you can get a gen 5 device it will interact with the host system at true gen 5 speeds even if the downstream devices are gen4. Very difficult to get in the consumer space but in theory this is the only way to get the full bandwidth of a gen 5 slot if using gen 4 endpoint devices (eg a 16 lane to 32 lane gen 5 switch will happily provide simultaneous full speed access to 8 gen 4 NVMe drives because the gen 4 drives only force the connection between the switch and the drive down to gen 4, the upstream link is separate and can run at gen 5 as long as the host can).
This kind of thing is why Intel needs to be in the GPU space. This is a big market for them and they're just getting started I think.
I think Furmark 3 is just going to be Epic Sax Gandalf except with realistic hair texture.
AMD also has this capability for ESXi, Citrix, and Linux (this is public information - MxGPU - and has existed for a long time). But the up to date tooling to enable it isn't outright available. 😔