"You just buy our hardware, you get it, and it works." What a world we live in, where "We won't scalp you for more money to use hardware you already paid for" is a positive selling point. Imagine if other industries were like that. "I know you spent a lot of money on that washer, but we won't let you wash pants in it unless you pay for the subscription."
That's pretty much always been the world. We have food purity laws for a reason. Prior to those laws companies would advertise their product being more pure than the others, and if you got a super cheap loaf of bread it might have a good bit of sawdust in it. Speaking of bread, back in ancient Roman times they used bronze bread stamps to identify the baker who baked the bread, so they could punish the correct person for selling adulterated food. As long as there's been profit there's been greed, and there's been people more than willing to fuck you over for a few extra pennies.
I was the project lead during the original pre-beta nVidia Grid program here at the college using HV 5 with K1 cards (VMWare, nVidia & Teradichi NDA's). Worked extremely well for what we wanted to do VDI-wise. When everything shifted to Grid 2.0 & Subscriptions only we killed all future accelerated VDI investment and maintained the K1's as long as was feasible - was just too costly for what we were trying to accomplish under subscriptions unless we increased tuitions, which we refused to do at that time (Zero student fees at that time). AMD's effort never seemed to work well so we skipped that. We are waiting to see how this Intel card shakes out to possibly retro-fit into all out NX-8155N-G8's (AOS/ESXi based) that had the GPU enablement kit installed, just without a GPU at purchase.
old AMD cards didn't have decent encoders nothing anywhere near as good as even the old Nvenc on K1, RDNA2-3 or Intel stuff, is about as good as Turing or newer, and AV1 is a thing these days too
@@XDbored1 The thing with AV1 is you need your VDI software to support it. Which, given its openness, you would expect everything and their mother already support it but...
Politicians: Hey! We are losing out on taxes here. Nvidia, we agree with your lobbyists and their briefcases and jackets with gold bars sown in. We must do something! For the people's safety, of course!
@Sickling3 seriously? Let's huff paint fums together, lol. Just like the political system in the USA, the only way it will change is if the population as a whole stops everything, going to their jobs, shopping, etc, for a week to show the establishment they really have no power once the peasants say no. We live in neofeudalism
Until about a year ago I worked at a rural bank that heavily used VDI to work around the horrendous 10-20 mbit internet connections we could get for most of our branches. Instead of clogging that pipe with Windows Updates, file transfers, remote troubleshooting, web traffic, etc. its just a single video stream with a bit of USB communication for the keyboard, mouse and receipt printer and provided an acceptably snappy experience for our workers at all locations, and most importantly avoided the challenge of roaming profiles making someone's first login at a remote branch take 10 minutes (we had several people who would frequently travel between branches) At the bank we specifically were running Citrix VDI on a cluster of 3x HP DL380 G9s with storage offloaded to 2x mirrored DL380P G9s for both file storage and iSCSI targets for the VM storage. No video acceleration, just raw CPU power. It was...passable. The neglect of Citrix was pretty clear, and the compression got annoying at times, but most importantly people could work from any branch and have basically the exact same experience and not notice any difference whether it was a branch with symmetric gigabit fiber or one with just a DSL connection for century old phone lines.
@@nokkonokko this type of setup is common in finance especially for users in regions where you may have restrictions on where you can store data or how you can store it.
@@Damien_N Correct. People may shit on the whole VDI thing and logging in to another computer from your own computer, but this is mostly because of security reasons and in the case of the original poster, could also be because of limited bandwidth and not wanting to deal with the whole Azure/Intune combo that is very prevalent nowadays. As much as I hate dealing with Citrix support and some of their products, the end result is still impressive. I have seen customers with a ~ 20 mbit connection with something like 15 people in the office + VoIP + some whatever camera solution that was streamed to some third party for archival and security reasons. It is a miracle that it even worked for all these years to begin with, but it was barely enough to do whatever they needed to do.
Old news, I first designed & implemented Citrix for a well-known global bank with international presence in 1997 for 24x7 trading access using the internet as the backbone, with Checkpoint Firewall-1 and ISDN as the backup! But 64kbps was plenty then ;-) When Y2K happened, I had the terrible chore of physically visiting all those sites dotted around the planet again to patch them. It wasn't all a bed of roses though, support calls in the middle of the night are no fun :-(
This is exactly what i want, I'm building a genoa server, but need hw acceleration for my vms, but i don't want to stuff my server full with graphics cards, I want one graphics card shared between average vms, and one topnotch graphics card for a vm where i do rendering and some other heavy lifting calculations. Finally a card with SRIOV without subscriptions and licensing I'm no Intel fan, but in this case glad to give them my money. Can't wait to see your video Wendel.
We are using a classic Windows Terminal Server without virtualisation. The experience with hardware acceleration by a GPU is great! Without GPU acceleration the work with CAD applications wouldn't even be possible.
(large public university) we run on-prem citrix daas for the engineering, gis, etc students and nvidia gpus have always been the bottleneck for us. Requiring even delimitations for vgpus makes it so we have to have some vlabs with less vram than they really should, or have less concurrency than we need. Without the nvidia subscription licensing we could probably afford enough servers to have concurrency and performance where they need to be. Very excided to see if I can get the beancounters to let me play with a Flex card.
Yas, this pizza box is so low cost, its worth buying one in this context for the R&D. You'll be pleased to hear I got everything up and running with proxmox 8.1 for the next video and it is AWESOME. Does require some minor not-intel-sanctioned hackery.
Two things: Firstly, as someone who messes with VDI in the homespace it would be fun and interesting to tinker with, but the cost right now is way to much for that. Will they offer a lower performance one that can hit the $300-$500 price point? Secondly, can they please make it so that if the Intel CPU detects an Intel ARC GPU, that it doesn't disable Intel Quicksync? If they could make the ARC and integrated GPU work together while keeping Quicksync capability I feel that would really boost the market for Intel GPUs.
See the comment from: butterscotchpanda There are guides out there for flashing Flex firmware onto A750 or A770 GPUs, since both use the DG2 die so it is technically possible.
Fascinating stuff. My employer (a German publishing company) is years behind on this. For remote work we have to keep our desktop PCs that are on premise running 24/7 so we can connect to them via RDP. 😄More than 100 desktop PCs and if anyone would be using "your" machine locally, you are SOL...)
Not a customer, but VMware victim here. Hoping the new broadcom bargains heavily punish my company for going all-in on virtualized slices of decade old servers as employee desktop replacements.
My former employer went in balls deep. So deep I don't think you could see the top of their head by the time I left. They have a massive, read MASSIVE, VmWare footprint. I wish them and Broadcom all the luck in the wars to come. I personally think Amazon and all the other cloud providers should also increase prices 10x. That's the current paradigm apparently. Kind of like the "F*ck you, pay me" scene in Goodfellas.
@@vincei4252 The cloud vendors are already 16 times more expensive than COLO for the same end user experience when I'm benchmaking my application stack.
I tried SR-IOV on a 14th generation Intel iGPU to hardware-accelerate VMs, and it works great. However, no free remote clients are as lag-free as Nvidia GameStream and Moonlight. This is definitely the future of computing. It's strange that Nvidia has stopped support and locks down their hardware. Some governments should investigate this forced hindrance of compute evolution and innovation, as well as the price manipulations and control exerted by a significant portion of the hardware industry today. The market has become somewhat dysfunctional, with too few competitors controlling and holding back changes that could have large efficiency benefits for the whole world, in exchange for the possibility of more profit.
Have you tried enabling Nested Virt alongside SRIOV for the same VM? Because since 12gen Intel seems like not possible with at least host-passthrough of CPU, and has to be host-model in order to enable that, but then the SRIOV of Intel iGPU gives error 43, like if it had to be the exact model, otherwise iGPU will refuse to run. But since nested virt will also refuse to run with the exact model, i can't use it alongside SR-IOV, which was my goal in te first place (to use these both at the same time). But seems weird, as i can use an ARC A770 with nested virt enabled (VFIO pass through, no SRIOV tho as it isn't compatible OC).
Hah! Glad I'm not the only person taping up vents to push more through passive cards. We had a load of late model Tesla accelerators we needed to test (ITAD) and they only kept workstations around for the admittedly-rare times we got PCIE cards. The only way to get enough airflow was to tape off a bunch of the exhausts and fix the fans at max! It worked though!
I kinda miss the days of tape and fans. Back in the day I had an Athlon 4 I OC'd hard enough the only way to keep it cool was to replace the side panel with cardboard that had a cutout for a 9" high velocity fan. Back when you could get meaningful uplift from pushing your hardware.
I had a really frustrating experience trying to get Dell to even talk to me about the DC Flex GPUs. I had to contact people at Intel and get them to talk to Dell and admit it exists. The Flex DC are very interesting from a video encoding perspective. Apparently they work well with Docker for that.
I worked in this space in the game streaming environment. It is incredibly complicated to build that streaming protocol and encoding. It's kept proprietary as it is market advantage for each of the respective platforms. Having it commoditzed will minimize platforms still attempting high performance video capture remotely as a service offered on the clouds.
In my dream world I would have one server in my basement and multiple wireless keyboard & mice with a screen that has a built-in VDI client. No computers in the house, only the input and output devices.
In theory you can circumvent some the VDI-client stuff by streaming the GPU accelerated apps and output to the screen you're viewing it on. Microsoft App-V, as dead as it is today, also had a client for Linux meaning you could just "stream" any app from a MS Windows (terminal) server on your Linux rig (not sure if it worked for MacOS) and be done with it. I don't recall if App-V supports GPU acceleration but I wouldn't be surprised if it did. Maybe there will be a replacement software stack to stream apps from an OS- and (CPU/GPU) architecture agnostic platform to whatever device you want. Would be cool.
Yes, or even further one computer in the neighborhood and during prime time you have 1x performance but off hour you have 10x. Not everything scale great today, but if we were to work this route there is a lot of potential benefits like shared memory, storage etc. Just thinking on how badly we utilize hardware and power today by multiple idle computers everywhere you could just grasp the effect this could have on the hole society in the future.
That is super cool! I built a 2 GPU (K80, M40) with an older SuperMicro server. Running a few VDIs for our marketing team. It was easier to keep the workstations near the storage server and have 10 GbE. (Running Proxmox) I would love to try out with better update path. Great use case when we have people calling out and they don’t have (company provided) computers at home.
So glad you mentioned taping off the slot, that was my first thought as with a single slot card its the easier path for air rather than through the GPU heatsink. Its one thing that annoys me on modern desktop cases too, dust filters everywhere there is a fan but then they leave big open gaps on the back making it more critical you get your fan curve right to avoid the GPU sucking in dust that way.
Maybe a stupid question, but could multiple VMs seamlessly use one GPU (not necessarily at the same time, but...)? I mean I might like to make a homelab setup where I can remote into one of several VMs for different purposes, without needing multiple GPUs, and without needing some janky setup to change GPU passthrough each time I remote into one. I'm not picking up if these Intel Flex GPUs will make this any more supported. edit: oh, you mentioned SR-IOV passthrough, maybe I gotta search this new buzzword to learn more...
Great video. We use GPUs in OpenStack. Works great. Intel XE graphics is pretty good in Linux, but it needs the backport drivers; however once the driver issues are overcome it is an all around fantastic solution, especially for VDI/video encoding/etc.
Keep working on it, Wendel. I’ve been working towards configuring my family’s VMs through proxmox and have been fiddling with Tesla p4 core sharing in a hpe proliant dl360 gen9 home server. I’m intrigued at the opportunities with the intel flex 140 and 170 options. My server only has a 75 watt tdp on the pci express gen 3 flex connection. The 140 would match the 75 watts, but would bottlenecked with the gen 3 pci. All the same, looking forward to see the fruits of your hard work! Thanks!
We are Vmware horizon shop. We've been running GPU since the very first one was available for horizon. For no apparent reason, we are going to be switching off of VMwhere as fast as possibwe are planning to do the whole migration to XCP - Ng, but the GPU pass through is the sticking point. We are extremely hopeful that Intel CPU's will allow us to do this. We have been waiting to get ahold of one for this exact use purpose. If not xcp-ng then attest hyper-v with gpu-p
Just out of curiosity: have you or your employer considered Nutanix? I don't know the pricing differences between XCP-NG, Hyper-V and them but from what I hear Nutanix is apparently the closest in terms of features and usability of VMware.
@masterTigress96 nutanix is more expensive, and much more limited on hardware support. Outside of vdi xcp-ng has everything we use vmware for and is cheaper than ether one. We have a way forward with horizon without a vmware back end, but it's not elegant to say the least. Plus we have always hated the Nvidia licensing they added, so if we can drop that for the cost of some new cards, we are all in.
Hi, I built one of the largest VDI deployments in the state of california in the public sector. We used VDI for many things, but always found the cost of graphics cards being included made the solutions value drop significantly. Adding millions to licensing and hardware costs wasn't a great way to move forward. Luckily in our use case, most work that was to be done was productivity solutions and internet browsing. That being said, there is a huge opportunity for people to have distributed VDI devices with this card, and Partners to build a practice of supporting those hypervisor solutions you mentioned. The other thing you didn't speak about immediately is density. When we architected our solution, we went with 185 Users per Server (supporting a total of 10K in our initial roll out). Can this card provide graphics capabilities without a limit to VDI or RDSH devices? that's the real question is how far will it scale.
I have a gtx 1060 6gb in my server, i use it for h265 video encoding in windows server 2019 I have also tried to activate remote fx for the remote desktop I use to connect to the server but i dont know if it works
Broadcom/VMware can go suck an Nvidia sized GPU. I have the use case for VDI, but the price advantage just is not there given the workload I have. Hardware wise, it would be, factoring in the software licenses, it comes out even. Edit: Have to do the math for an Intel Flex & Proxmox based system tomorrow...
I manage a medium manufacturing plant and inherited some overkill high availability cluster setup that was never actually implemented. We don't use VDI but with aging desktops and more emphasis on engineering/R&D I've been tossing around the idea of something like this. I was just considering testing out a Tesla P40 as most of our CAD work is pretty basic, but needs hardware acceleration.
Now we want that Flex 170 vBIOS SR-IOV compatibility sticked and implemented to the ARC A770 16GB for those we aren't businesses (as they don't have a licensing model per user why worrying about only businesses) and/or we are just poor enough enthusiast customer/businesses working along side KVM-HyperV nested virtualization to be able to run WSA/WSL2 dev environments on a modest VDI homelab infraestructure. Nowadays only consumer cards compatible with that GPU virtualization are Intel iGPUs (whether if through SR-IOV, and or Intel GVT-g) and/or RDNA2/A770 through VFIO (but since this last one depends on Resizable BAR for good gaming perf, and that's difficult to pass to VM, I am not currently using it). I wanted to believe that RDNA3 would be a good substitute for RDNA2, but after testing with Navi 31 and Navi 32, those cards will refuse to output any video signal at all if nested virt was enabled, regardless if it was sticked onto an AMD or Intel platform. I'm not mad, just dissapointed with Radeon team after RDNA2 successful lineup.
I've been running a remote desktop system for the last few years, locally. I have a small machine shop business (it's just me), and all of my desktops are running linux, but the CAD/CAM software I use is windows only. I started out with a windows 10 VM in proxmox, and it was pretty alright. I had a GPU passed through directly to that VM, and I accessed it via parsec, but for whatever reason, there was still some lag that was a little annoying. I THINK it was due to the virtualization layer, and not an encoding problem. Currently running the same setup, but with windows installed bare metal on a newer system in my rack, still accessing via parsec, but I'm using the intel onboard graphics for encoding instead of a GPU. Been working pretty great, but my CAD/CAM system doesn't need any GPU, so that may change if I move do a different software stack.
We used VMWare and Windows 2012 r2 with RDP on Windows 7 and 10. Once set up it was so safe and easy to administer. The company would not fund GPUs. It was a full SQL application.
So I'm pretty new to all of this type of stuff but I think that I have a use case where this would really work well: I'm a grad student in a MS of Electrical Engineering program. On the project that I am on we sometimes use ArcGIS Pro. We have a machine with a nvidia RTX3080 GPU in it that we can remote into and run models on. The catch is that if it's a large project, in order to "pan around" or to examine say a 3D terrain model we have to actually save the file/model, then sync the files to our local machines, then open ArcGIS on our local machines and check things out. If we try to pan around (sort of like what the guy from intel was doing in google earth) its very stuttery/glitchy. Am I correct in assuming that this would mean that instead of the graphics for what is being shown on the screen being rendered on the CPU it would be rendered on the GPU so it wouldn't be stuttery? Another thing that has always annoyed me but I'm not sure if it has any bearing on this at all or not: I remote into a Linux machine that is right next to my desk (just easier than swapping keyboard/mouse/monitors around, and yes I know KVM's exist but I have two 120 Hz ultrawides, and KVMs that support 2x 5120x1440 @ 120Hz are VERY expensive). Anyway the main reason I do this is because the python library tensorflow does not support using GPU's for training in Windows. I have some programs (like ArcGIS Pro) that only support Windows so I have to run it. Anyways, the two computers are connected via a 2.5 Gig ethernet link and the Linux machine has 2x RTX4090's in it. When I scroll around in a program in Ubuntu (via remote desktop) it sometimes can be glitchy just like how ArcGIS can be glitchy when panning around on that other machine. I also don't know how to explain it but the "look" of the Linux desktop is just not as "clean" as when viewed directly. Maybe something to do with aliasing? The machine that I am using to remote in is running W11 and has an RTX 3080 so I doubt it is on the receiving end.
I want to see this for all the annoying edge applications (Security video workstations / command centers, I'm looking at you!) being in one expensive box instead of spread all over the world. Then I can secure VPN in the clients that need access to the network and there are no ports on the actual network that allow access. Same goes for the industrial control / lighting control / MEP control servers. If you can create a great remote desktop experience we could start walking back vendors wanting a VLAN they have L2 access to throughout the building a bit.
I need confirm, but i'm fairly certain that the lack of hardware acceleration is what has held back our VDI hopes and dreams @ work for years. This seems like something that could push it out of the PoC phase and into at least a pilot.
I remember hearing about this tech nearly 11 years ago and was so amped for it to take over, then all the companies ruined it. Honestly reminds me of containerization... everyone was using it and had it but no one wanted to share until Docker almost went under and open sources to stay alive.
I thought you could just go and buy the GPU Flex... But they are 1300€ and 1800€, respectively. Which is maybe nice for the novelty of having a dual GPU card FLEX140. (Intel had a line of products called server GPU, which used 3 or 4 laptop CPUs with just the iGPU enabled for transcoding). Intel even has Windows drivers for the FLEX GPU product. It'd just like an A770 mainly.
As a user with a homelab who deals with a bunch of used enterprise VDi capable Gear & some gaming shenanigans, I am frustrated by subscription models and I often have to tackle that even as a simple hobbyist. NOT having a subscription is VERY appealing. The real special sauce is... CAN we use this hardware in proxmox? Perhaps eventually.
I'm running GPUs with passthrough on my current Proxmox cluster but not for VDI, though I've been interested in trying it. Something like this really piques my interest!
I have VMs with GPUs I can use for gaming and video editing. The problem is all the remote access programs I have tried suck and have too much latency. Also wish I could easily virtualize GPU resources. Granted I am still running esxi...
This seems really cool, we do small clusters for clients with remote desktop and i wonder if this may increase performance for the end user. Would be interesting to test once the software is there for proxmox :D
So not for enterprise but I do use a VM for my personal machine. It is nicer because in many ways I can access desktop hardware on any cheap laptop. 15 year old macbook air running linux can be a client and run fine. I personally use truenas scale instead of proxmox because the server I am doing this on is primarily a storage server but it works alright.
Over the last year we have migrated a lot of servers to SAAS or into Azure which dropped out vmware count. By our renewal cost though you wouldn't know it. Our VMware hosts are coming up on age and will either be replaced by something else(prob hyper-v) or we might just move the remaining stuff to the cloud. We renewed VMware for this year but I don't see us renewing past next year.
Yet another video showcasing the usefulness of a full height GPU in a server, but I'm stuck with my 2U with only vertical slots (so half-height). I wish I could find a generic adapter to convert some slots to horizontal full-height but that magic seems relegated to only servers with proprietary motherboards.
Is there Bifurcation on that slot? That would be nice to rig something up so the GPU can still be able to use another 8 lines for something else. Only thing that will stop that is the 1u format of the case. Shame it looks like a very good servers.
Question how well could this work as a media or streaming server and would this technologies go downstream towards the more prosumer side of things? Intel has been doing good work with their video encoding and Im thinking of building a pc for that. It seems right now the use case for this is remote desktops which is great for enterprise but loved to see how this technology works for more local uses.
RDP hw accel is essentially just the same thing as live media transcoding, so it'll work as good if not better because it's not heavily time dependent. I'd imagine you could even do both at once through a little bit of shenanigans. Ultimately you just need the drivers.
Yeah... We've been VERY CAUTIOUS/concerned about the broadcom buyout. RHEL Openshift is on our roadmap so we're slowly flexing up to shift platforms away from vmware to something else. (we just signed the renewal in December so we're locked for another 2 years safely at the moment).
I do VDI for Personal Use with 2 Tesla P4s used to have consumer Cards, but those are just the sweet spot for light compute and gaming. I use linux on my laptop but wanna Game sometimes on the go thus I have a couple Windows VMs for me and friends and a Plex using the 2 P4s
I was looking for this comment! After seeing your other videos on VDI, I immediately thought of you! Really excited to see what/if you do anything with this!
Well one thing that came out of the broadcom deal, it is finally forcing my organization to get rid of the asinine (imo) deployment method we had. We had a bunch of small servers that each had VMware on them, with a single VM each. As the DevOps guy, they didn't let me touch anything even at the OS level let alone hypervisor. But now they're just going for a proper coreos pxie boot setup and both the performance is better and my program isn't going to have to pay the infrastructure team for the convenience of their asinine decisions.
I bought a tesla m10 to try this out myself a while ago, I had no idea there was licensing involved, hopefully that's nothing I need to pay for to use that card, or I'll resell it, and I planned on using an a770 in my machine to try av1 streaming anyway
What do you think of using this in a production studio to have a VM for each Remote editor? Have an office NAS. Each VM 10GbE to a NAS. Remote editors run editing software on VM and remote edit through their remote connection to the VM. Would this work a lot better than using Parsec for remote editing?
Work with a company that pays a lot for Nvidia license on hardware acceleration for remote desktops for engineers. The whole VMWare/Nvidia licensing is pushing for a change and this is something to consider. This would be great for PCI DSS compliant environments as well they are best served by remote desktop system.
Another VmWare victim here. Will watch while Broadcom is cutting VmWare apart TCM-style. I'm not convinced they won't even let the virtualisation products die a slow and agonizing death. Also looking forward to the class actions once Broadcom inevitably flips all the vSphere-customers the bird ;)
There are going to be fireworks that's for sure. My distributor is a major licensing disty in the UK, and broadcom just killed their partner status. Just out of nowhere.
I use a pair of Tesla P4s within my Proxmox server to provide GPU accelerated gaming VMs to some of my friends. I would love to have a better solution, with proper cooling, and without unnecessary licensing... The whole vgpu drivers thing paired with that they don't even have their own VDI solution, and the streaming part has to be figured out manually is just a joke. I am gonna buy an Intel GPU as soon as it has at least Pascal level of performance and supports SR-IOV without licensing/subscription/whatever... Especially since Nvidia drops Pascal series Tesla card support from GRID drivers starting from version 17...
Removing that Subscription is a Blessing . many times i cannot keep the services paid and need to be able to still acccess services , its only at certan times as budget allows but still down time
I understand the security need for protecting IP, but even with HW acceleration, you are requiring users to remote in to a clean environment with locked down permissions just to then SSH (through putty....) to a head node. Slows the pace of prototyping dramatically
So as a sysadmin that has only ever deployed classic stile RDS solutions on Hyper-V and VMWare. Does this mean anything? Can I stick in an intel flex GPU in ESXI or Hyper-V hosts and utilize them in RDSH?
Not enterprise, but at home I have a EPYC CPU running esxi8 and have a dedicated "Steam Gaming VM" that passes through a 4060 that I stream to my handheld devices. It works perfectly but would love to be able to share it with my other VM's on the hypervisor. Unfortunately PCI only works with one VM only but I'd be open to installing a second if VMware allows for it to share its resources to any VM that requests it. I'd love to know the noise level of this Pizza Server. Great video btw
This isn't just about VMWare though, it's specifically about SR-IOV which is a hardware feature that allows much better graphics acceleration by giving each VM its own chunk of real GPU, rather than translating and moving rendering calls in and out of the guest with a ton of extra overhead to prevent VM escape
This is a super cool stuff! Please don't just focus on Windows clients though, Linux should get some validation love, too; Also perfectly on time, as I was thinking about building some VDI infra as our company is growing. I think I still have 6-8 months until I need this, but it would be great to use the Intel Flex and Proxmox combo.
Really wish someone would make a card that only did media acceleration. I have a Quadro P2000 for Plex and the drivers under Ubuntu just mysteriously crap out and disappear every few months. I spend a whole day trying to get it working again. Was looking at the ARC GPUs and the better ones need PCIe power. Well, my SuperMicro 3U chassis doesn't have any spare power to run a GPU. I don't need all the horsepower a GPU has to offer just to be wasted for media decoding and encoding.
Would this work as a "Workstation" server for a small office/home? Could I install one of these at the house and then everybody in the house simply have their own remote desktop instance?
Yes, but it would be bad value for money for that purpose if the specific appliance being discussed were used, 5000 dollars buys a lot of clients and the Flex 170 isn't actually that powerful in terms of raw rasterisation
@@bosstowndynamics5488 I was more referring to the GPU in the video. I'd actually like to re-purpose my Threadripper Pro, which I'm currently using (and entirely underutilizing) as my workstation. I'd like to build some kind of machine that can host about 5-6 remote desktop instances that have virtually zero video latency (on the LAN).
@@jeremyrangel8138 Yes but again, the expense. The Flex 170 is an extremely expensive GPU (a quick check of eBay found a single example listed for nearly 2800 Euro, which is more than half of the cost of this entire system). And remember, the 170 is really just an Arc a770 with the GPU equivalent of virtualisation extensions enabled, so it's not even a particularly powerful GPU. 5-6 is right on the edge but I'd be willing to bet that you'd be slightly better off in terms of price to performance by running 6 low end dGPUs and dealing with the slightly more complex wiring/a big case with some Oculink adaptors/PCIe breakouts (6 GTX 1060s could be had for a lot cheaper and would easily run circles around 1/6th of a Flex 170 each even with only a few lanes of PCIe to play with). You could pay for convenience here but this specific server is probably going to be close to the cheapest way to get a Flex 170 build up and running.
@@jeremyrangel8138 Not sure if the other comment is showing but with a TR Pro on hand already your easiest route is to find a decent example of a single slot GPU and slap 6 of them in, you've got plenty of PCIe to spare and 6 cheap GPUs will both cost less than, and run a lot faster per user than, a Flex 170 sliced up 6 ways. The 170 only really makes sense if you want to provide a little bit of GPU to each of 10+ clients
@@bosstowndynamics5488 Are you talking about connecting each user's monitor/peripherals via cable? If so, that's not really feasible. The farthest user from the rack would be about a 100ft cable run....
So glad Wendel got away from Teksyndicate and started a real tech channel.
Same. I totally forgot whats his face that used to run it with him, he was awful. Wendel really carried the show, he has a lot of charm.
Did i miss any news?
@@Splarkszter this is years ago before level1
Wait what, I didn’t know that
@@Splarkszter Wendell used to be Teksyndacate's magic elf and did videos for/with him. He eventually left and made his own thing
"You just buy our hardware, you get it, and it works." What a world we live in, where "We won't scalp you for more money to use hardware you already paid for" is a positive selling point. Imagine if other industries were like that. "I know you spent a lot of money on that washer, but we won't let you wash pants in it unless you pay for the subscription."
I am afraid it might become a really for everything soon. We are at the end of the era of ownership.
There are condos with coin op laundry. People are paying to wash their pants at home!
To insure user safety, this washer only washes genuine Maytag pants...
That's pretty much always been the world. We have food purity laws for a reason. Prior to those laws companies would advertise their product being more pure than the others, and if you got a super cheap loaf of bread it might have a good bit of sawdust in it. Speaking of bread, back in ancient Roman times they used bronze bread stamps to identify the baker who baked the bread, so they could punish the correct person for selling adulterated food. As long as there's been profit there's been greed, and there's been people more than willing to fuck you over for a few extra pennies.
@@vezquexMentally chall*nged?
I was the project lead during the original pre-beta nVidia Grid program here at the college using HV 5 with K1 cards (VMWare, nVidia & Teradichi NDA's). Worked extremely well for what we wanted to do VDI-wise. When everything shifted to Grid 2.0 & Subscriptions only we killed all future accelerated VDI investment and maintained the K1's as long as was feasible - was just too costly for what we were trying to accomplish under subscriptions unless we increased tuitions, which we refused to do at that time (Zero student fees at that time). AMD's effort never seemed to work well so we skipped that. We are waiting to see how this Intel card shakes out to possibly retro-fit into all out NX-8155N-G8's (AOS/ESXi based) that had the GPU enablement kit installed, just without a GPU at purchase.
Yes, the grid subscription was a farce unfortunately. HV5 was a cool era, we all thought that vdi was the future of endpoint computing, right?
old AMD cards didn't have decent encoders nothing anywhere near as good as even the old Nvenc on K1, RDNA2-3 or Intel stuff, is about as good as Turing or newer, and AV1 is a thing these days too
@@XDbored1 The thing with AV1 is you need your VDI software to support it. Which, given its openness, you would expect everything and their mother already support it but...
No per user license and no subscription. Hear that Nvidia! No double milking the customer
An unheard of strategy by today's, everything is a subscription based IT business model 😮
Politicians: Hey! We are losing out on taxes here. Nvidia, we agree with your lobbyists and their briefcases and jackets with gold bars sown in. We must do something! For the people's safety, of course!
i hope broadcom quits doing the subscription bullshit after purchasing vmware.
@Sickling3 seriously? Let's huff paint fums together, lol.
Just like the political system in the USA, the only way it will change is if the population as a whole stops everything, going to their jobs, shopping, etc, for a week to show the establishment they really have no power once the peasants say no. We live in neofeudalism
@@Sickling3 Unfortunately Broadcom ruins VMware right after the acquisition
Until about a year ago I worked at a rural bank that heavily used VDI to work around the horrendous 10-20 mbit internet connections we could get for most of our branches. Instead of clogging that pipe with Windows Updates, file transfers, remote troubleshooting, web traffic, etc. its just a single video stream with a bit of USB communication for the keyboard, mouse and receipt printer and provided an acceptably snappy experience for our workers at all locations, and most importantly avoided the challenge of roaming profiles making someone's first login at a remote branch take 10 minutes (we had several people who would frequently travel between branches)
At the bank we specifically were running Citrix VDI on a cluster of 3x HP DL380 G9s with storage offloaded to 2x mirrored DL380P G9s for both file storage and iSCSI targets for the VM storage. No video acceleration, just raw CPU power. It was...passable. The neglect of Citrix was pretty clear, and the compression got annoying at times, but most importantly people could work from any branch and have basically the exact same experience and not notice any difference whether it was a branch with symmetric gigabit fiber or one with just a DSL connection for century old phone lines.
…mainframe computing in the 2020s…
@@nokkonokko this type of setup is common in finance especially for users in regions where you may have restrictions on where you can store data or how you can store it.
@@Damien_N Correct. People may shit on the whole VDI thing and logging in to another computer from your own computer, but this is mostly because of security reasons and in the case of the original poster, could also be because of limited bandwidth and not wanting to deal with the whole Azure/Intune combo that is very prevalent nowadays. As much as I hate dealing with Citrix support and some of their products, the end result is still impressive. I have seen customers with a ~ 20 mbit connection with something like 15 people in the office + VoIP + some whatever camera solution that was streamed to some third party for archival and security reasons. It is a miracle that it even worked for all these years to begin with, but it was barely enough to do whatever they needed to do.
Old news, I first designed & implemented Citrix for a well-known global bank with international presence in 1997 for 24x7 trading access using the internet as the backbone, with Checkpoint Firewall-1 and ISDN as the backup! But 64kbps was plenty then ;-)
When Y2K happened, I had the terrible chore of physically visiting all those sites dotted around the planet again to patch them.
It wasn't all a bed of roses though, support calls in the middle of the night are no fun :-(
As a tech for a VMware customer we are "looking to diversify our virtualization environment."
This is exactly what i want, I'm building a genoa server, but need hw acceleration for my vms, but i don't want to stuff my server full with graphics cards, I want one graphics card shared between average vms, and one topnotch graphics card for a vm where i do rendering and some other heavy lifting calculations. Finally a card with SRIOV without subscriptions and licensing I'm no Intel fan, but in this case glad to give them my money. Can't wait to see your video Wendel.
We are using a classic Windows Terminal Server without virtualisation.
The experience with hardware acceleration by a GPU is great!
Without GPU acceleration the work with CAD applications wouldn't even be possible.
(large public university) we run on-prem citrix daas for the engineering, gis, etc students and nvidia gpus have always been the bottleneck for us. Requiring even delimitations for vgpus makes it so we have to have some vlabs with less vram than they really should, or have less concurrency than we need. Without the nvidia subscription licensing we could probably afford enough servers to have concurrency and performance where they need to be. Very excided to see if I can get the beancounters to let me play with a Flex card.
Yas, this pizza box is so low cost, its worth buying one in this context for the R&D. You'll be pleased to hear I got everything up and running with proxmox 8.1 for the next video and it is AWESOME. Does require some minor not-intel-sanctioned hackery.
Oh man 15 clients running at 4K 60... Sign me up!
420k subs. Blaze!
Huge congrats to Level1Techs!
Cheers!
Hell yeah, finally.
Very excited for the proxmox implementation video
Two things:
Firstly, as someone who messes with VDI in the homespace it would be fun and interesting to tinker with, but the cost right now is way to much for that. Will they offer a lower performance one that can hit the $300-$500 price point?
Secondly, can they please make it so that if the Intel CPU detects an Intel ARC GPU, that it doesn't disable Intel Quicksync? If they could make the ARC and integrated GPU work together while keeping Quicksync capability I feel that would really boost the market for Intel GPUs.
See the comment from: butterscotchpanda
There are guides out there for flashing Flex firmware onto A750 or A770 GPUs, since both use the DG2 die so it is technically possible.
I want the true pizzabox to come back. Stuff like the Mac LC style, where it was almost literally the same size.
Fascinating stuff. My employer (a German publishing company) is years behind on this. For remote work we have to keep our desktop PCs that are on premise running 24/7 so we can connect to them via RDP. 😄More than 100 desktop PCs and if anyone would be using "your" machine locally, you are SOL...)
Welcome to germany. "Das Internet ist für uns alle Neuland"
Flex drivers for tha a770 would be awesome 👍
There are guides out there for flashing Flex firmware onto A750 or A770 GPUs, since both use the DG2 die so it is technically possible.
@@butterscotchpanda where?
Looking forward to that little teaser, Wendel. Have been wondering when that would arrive.
Not a customer, but VMware victim here. Hoping the new broadcom bargains heavily punish my company for going all-in on virtualized slices of decade old servers as employee desktop replacements.
My former employer went in balls deep. So deep I don't think you could see the top of their head by the time I left. They have a massive, read MASSIVE, VmWare footprint. I wish them and Broadcom all the luck in the wars to come. I personally think Amazon and all the other cloud providers should also increase prices 10x. That's the current paradigm apparently. Kind of like the "F*ck you, pay me" scene in Goodfellas.
They want virtualization because they can spy on you, in real time if nessisary.
@@vincei4252 if they do that, Broadcom will x100 prices and scream the same but louder
@@vincei4252 The cloud vendors are already 16 times more expensive than COLO for the same end user experience when I'm benchmaking my application stack.
Proxmox at least is still an option
My body is ready for proxmox + sriov flex combo
It will be sweet when these hit the secondhand market. I would love one for my homelab
Wendell, whenever there is something that encodes Videos "like a beast" i'm all ears. I humbly request full coverage on that front
I tried SR-IOV on a 14th generation Intel iGPU to hardware-accelerate VMs, and it works great. However, no free remote clients are as lag-free as Nvidia GameStream and Moonlight. This is definitely the future of computing. It's strange that Nvidia has stopped support and locks down their hardware. Some governments should investigate this forced hindrance of compute evolution and innovation, as well as the price manipulations and control exerted by a significant portion of the hardware industry today. The market has become somewhat dysfunctional, with too few competitors controlling and holding back changes that could have large efficiency benefits for the whole world, in exchange for the possibility of more profit.
Sunshine
Have you tried enabling Nested Virt alongside SRIOV for the same VM? Because since 12gen Intel seems like not possible with at least host-passthrough of CPU, and has to be host-model in order to enable that, but then the SRIOV of Intel iGPU gives error 43, like if it had to be the exact model, otherwise iGPU will refuse to run. But since nested virt will also refuse to run with the exact model, i can't use it alongside SR-IOV, which was my goal in te first place (to use these both at the same time).
But seems weird, as i can use an ARC A770 with nested virt enabled (VFIO pass through, no SRIOV tho as it isn't compatible OC).
@@Pacho18 not tried with nvidia but with the intel igpu its not close to gamestream in my experience.
@@Pacho18 not tried nested virt
@@xpeditionsweden Do it. Is useful for deploying WSA/WSL2 dev/testing environments.
Hah! Glad I'm not the only person taping up vents to push more through passive cards. We had a load of late model Tesla accelerators we needed to test (ITAD) and they only kept workstations around for the admittedly-rare times we got PCIE cards. The only way to get enough airflow was to tape off a bunch of the exhausts and fix the fans at max! It worked though!
I kinda miss the days of tape and fans. Back in the day I had an Athlon 4 I OC'd hard enough the only way to keep it cool was to replace the side panel with cardboard that had a cutout for a 9" high velocity fan. Back when you could get meaningful uplift from pushing your hardware.
A nice surprise. I didn't expect to be excited about anything Intel got up to lately.
I had a really frustrating experience trying to get Dell to even talk to me about the DC Flex GPUs. I had to contact people at Intel and get them to talk to Dell and admit it exists.
The Flex DC are very interesting from a video encoding perspective. Apparently they work well with Docker for that.
I worked in this space in the game streaming environment. It is incredibly complicated to build that streaming protocol and encoding. It's kept proprietary as it is market advantage for each of the respective platforms. Having it commoditzed will minimize platforms still attempting high performance video capture remotely as a service offered on the clouds.
The dude on the right at 13:28 has some wild glasses
I remember the times when we used PCoIP offloading cards for our vdi solution. 🙂
In my dream world I would have one server in my basement and multiple wireless keyboard & mice with a screen that has a built-in VDI client. No computers in the house, only the input and output devices.
Sun Microsystems has entered the chat
@@Akkbar21 PDP-11 has entered the chat 🤣🤣
In theory you can circumvent some the VDI-client stuff by streaming the GPU accelerated apps and output to the screen you're viewing it on. Microsoft App-V, as dead as it is today, also had a client for Linux meaning you could just "stream" any app from a MS Windows (terminal) server on your Linux rig (not sure if it worked for MacOS) and be done with it.
I don't recall if App-V supports GPU acceleration but I wouldn't be surprised if it did. Maybe there will be a replacement software stack to stream apps from an OS- and (CPU/GPU) architecture agnostic platform to whatever device you want. Would be cool.
Yes, or even further one computer in the neighborhood and during prime time you have 1x performance but off hour you have 10x. Not everything scale great today, but if we were to work this route there is a lot of potential benefits like shared memory, storage etc. Just thinking on how badly we utilize hardware and power today by multiple idle computers everywhere you could just grasp the effect this could have on the hole society in the future.
Gattuso Fabric
Pepperoni Cheese Interface Express x16
Gluten Inside
Dish Delivered Rapidly 5
Those are some delicious puns
That is super cool! I built a 2 GPU (K80, M40) with an older SuperMicro server. Running a few VDIs for our marketing team. It was easier to keep the workstations near the storage server and have 10 GbE. (Running Proxmox)
I would love to try out with better update path.
Great use case when we have people calling out and they don’t have (company provided) computers at home.
So glad you mentioned taping off the slot, that was my first thought as with a single slot card its the easier path for air rather than through the GPU heatsink.
Its one thing that annoys me on modern desktop cases too, dust filters everywhere there is a fan but then they leave big open gaps on the back making it more critical you get your fan curve right to avoid the GPU sucking in dust that way.
Come on Matrox, the door is open to you again!
I'm ready too see it run proxmox!
Maybe a stupid question, but could multiple VMs seamlessly use one GPU (not necessarily at the same time, but...)? I mean I might like to make a homelab setup where I can remote into one of several VMs for different purposes, without needing multiple GPUs, and without needing some janky setup to change GPU passthrough each time I remote into one. I'm not picking up if these Intel Flex GPUs will make this any more supported. edit: oh, you mentioned SR-IOV passthrough, maybe I gotta search this new buzzword to learn more...
The Flex 170s do this out of the box, but arc a770 won't do it the same way. Intel igpu will do it too. Sr-iov support is what does it
Engagement Challenge! Really cool video, thanks for the interviews.
I do not see the Flex 170 listed anywhere on the store page. Are you certain that it comes with?
Great video. We use GPUs in OpenStack. Works great. Intel XE graphics is pretty good in Linux, but it needs the backport drivers; however once the driver issues are overcome it is an all around fantastic solution, especially for VDI/video encoding/etc.
Keep working on it, Wendel. I’ve been working towards configuring my family’s VMs through proxmox and have been fiddling with Tesla p4 core sharing in a hpe proliant dl360 gen9 home server. I’m intrigued at the opportunities with the intel flex 140 and 170 options. My server only has a 75 watt tdp on the pci express gen 3 flex connection. The 140 would match the 75 watts, but would bottlenecked with the gen 3 pci. All the same, looking forward to see the fruits of your hard work! Thanks!
We are Vmware horizon shop. We've been running GPU since the very first one was available for horizon. For no apparent reason, we are going to be switching off of VMwhere as fast as possibwe are planning to do the whole migration to XCP - Ng, but the GPU pass through is the sticking point. We are extremely hopeful that Intel CPU's will allow us to do this. We have been waiting to get ahold of one for this exact use purpose. If not xcp-ng then attest hyper-v with gpu-p
Just out of curiosity: have you or your employer considered Nutanix? I don't know the pricing differences between XCP-NG, Hyper-V and them but from what I hear Nutanix is apparently the closest in terms of features and usability of VMware.
@masterTigress96 nutanix is more expensive, and much more limited on hardware support. Outside of vdi xcp-ng has everything we use vmware for and is cheaper than ether one. We have a way forward with horizon without a vmware back end, but it's not elegant to say the least. Plus we have always hated the Nvidia licensing they added, so if we can drop that for the cost of some new cards, we are all in.
@@gigogigtub Interesting, thank you for the insight.
yea we use it, works fine in proxmox for years with novidya cards and amd as well.
Hi, I built one of the largest VDI deployments in the state of california in the public sector. We used VDI for many things, but always found the cost of graphics cards being included made the solutions value drop significantly. Adding millions to licensing and hardware costs wasn't a great way to move forward. Luckily in our use case, most work that was to be done was productivity solutions and internet browsing. That being said, there is a huge opportunity for people to have distributed VDI devices with this card, and Partners to build a practice of supporting those hypervisor solutions you mentioned. The other thing you didn't speak about immediately is density. When we architected our solution, we went with 185 Users per Server (supporting a total of 10K in our initial roll out). Can this card provide graphics capabilities without a limit to VDI or RDSH devices? that's the real question is how far will it scale.
I actually really love my A770 so I'm looking forward to that video
I have a gtx 1060 6gb in my server, i use it for h265 video encoding in windows server 2019
I have also tried to activate remote fx for the remote desktop I use to connect to the server but i dont know if it works
Broadcom/VMware can go suck an Nvidia sized GPU. I have the use case for VDI, but the price advantage just is not there given the workload I have. Hardware wise, it would be, factoring in the software licenses, it comes out even.
Edit: Have to do the math for an Intel Flex & Proxmox based system tomorrow...
I manage a medium manufacturing plant and inherited some overkill high availability cluster setup that was never actually implemented. We don't use VDI but with aging desktops and more emphasis on engineering/R&D I've been tossing around the idea of something like this. I was just considering testing out a Tesla P40 as most of our CAD work is pretty basic, but needs hardware acceleration.
Yea they're cheap used and work great. Lots of tools online to open it up to virtual gpus etc
Now we want that Flex 170 vBIOS SR-IOV compatibility sticked and implemented to the ARC A770 16GB for those we aren't businesses (as they don't have a licensing model per user why worrying about only businesses) and/or we are just poor enough enthusiast customer/businesses working along side KVM-HyperV nested virtualization to be able to run WSA/WSL2 dev environments on a modest VDI homelab infraestructure.
Nowadays only consumer cards compatible with that GPU virtualization are Intel iGPUs (whether if through SR-IOV, and or Intel GVT-g) and/or RDNA2/A770 through VFIO (but since this last one depends on Resizable BAR for good gaming perf, and that's difficult to pass to VM, I am not currently using it).
I wanted to believe that RDNA3 would be a good substitute for RDNA2, but after testing with Navi 31 and Navi 32, those cards will refuse to output any video signal at all if nested virt was enabled, regardless if it was sticked onto an AMD or Intel platform. I'm not mad, just dissapointed with Radeon team after RDNA2 successful lineup.
Exciting times… 👍🏾❤️
I've been running a remote desktop system for the last few years, locally. I have a small machine shop business (it's just me), and all of my desktops are running linux, but the CAD/CAM software I use is windows only.
I started out with a windows 10 VM in proxmox, and it was pretty alright. I had a GPU passed through directly to that VM, and I accessed it via parsec, but for whatever reason, there was still some lag that was a little annoying. I THINK it was due to the virtualization layer, and not an encoding problem.
Currently running the same setup, but with windows installed bare metal on a newer system in my rack, still accessing via parsec, but I'm using the intel onboard graphics for encoding instead of a GPU. Been working pretty great, but my CAD/CAM system doesn't need any GPU, so that may change if I move do a different software stack.
24 1080p 60fps 4k streams? Like, upsampled?
Nope.
We used VMWare and Windows 2012 r2 with RDP on Windows 7 and 10. Once set up it was so safe and easy to administer. The company would not fund GPUs. It was a full SQL application.
Color me impressed. I can't wait for the A770 video.
So I'm pretty new to all of this type of stuff but I think that I have a use case where this would really work well: I'm a grad student in a MS of Electrical Engineering program. On the project that I am on we sometimes use ArcGIS Pro. We have a machine with a nvidia RTX3080 GPU in it that we can remote into and run models on. The catch is that if it's a large project, in order to "pan around" or to examine say a 3D terrain model we have to actually save the file/model, then sync the files to our local machines, then open ArcGIS on our local machines and check things out. If we try to pan around (sort of like what the guy from intel was doing in google earth) its very stuttery/glitchy.
Am I correct in assuming that this would mean that instead of the graphics for what is being shown on the screen being rendered on the CPU it would be rendered on the GPU so it wouldn't be stuttery?
Another thing that has always annoyed me but I'm not sure if it has any bearing on this at all or not: I remote into a Linux machine that is right next to my desk (just easier than swapping keyboard/mouse/monitors around, and yes I know KVM's exist but I have two 120 Hz ultrawides, and KVMs that support 2x 5120x1440 @ 120Hz are VERY expensive). Anyway the main reason I do this is because the python library tensorflow does not support using GPU's for training in Windows. I have some programs (like ArcGIS Pro) that only support Windows so I have to run it. Anyways, the two computers are connected via a 2.5 Gig ethernet link and the Linux machine has 2x RTX4090's in it. When I scroll around in a program in Ubuntu (via remote desktop) it sometimes can be glitchy just like how ArcGIS can be glitchy when panning around on that other machine. I also don't know how to explain it but the "look" of the Linux desktop is just not as "clean" as when viewed directly. Maybe something to do with aliasing? The machine that I am using to remote in is running W11 and has an RTX 3080 so I doubt it is on the receiving end.
Do your screens only have one input? I have a similar setup but I just change inputs and have a kmv just for the k and m.
I really want to get hand on Intel FLEX for VDI stuff.
I want to see this for all the annoying edge applications (Security video workstations / command centers, I'm looking at you!) being in one expensive box instead of spread all over the world. Then I can secure VPN in the clients that need access to the network and there are no ports on the actual network that allow access. Same goes for the industrial control / lighting control / MEP control servers. If you can create a great remote desktop experience we could start walking back vendors wanting a VLAN they have L2 access to throughout the building a bit.
I need confirm, but i'm fairly certain that the lack of hardware acceleration is what has held back our VDI hopes and dreams @ work for years. This seems like something that could push it out of the PoC phase and into at least a pilot.
I remember hearing about this tech nearly 11 years ago and was so amped for it to take over, then all the companies ruined it. Honestly reminds me of containerization... everyone was using it and had it but no one wanted to share until Docker almost went under and open sources to stay alive.
Incredible 👍
VDI needs to bloom everywhere. And with ProxMox and XCPng not ESXi. And without Nvidia subscription.
Hopefully we'll see some used ones, in a few years for cheap.
This sounds amazing paired up with Proxmox!
I thought you could just go and buy the GPU Flex... But they are 1300€ and 1800€, respectively. Which is maybe nice for the novelty of having a dual GPU card FLEX140. (Intel had a line of products called server GPU, which used 3 or 4 laptop CPUs with just the iGPU enabled for transcoding).
Intel even has Windows drivers for the FLEX GPU product. It'd just like an A770 mainly.
Please let this work with TrueNAS. I'd love to use GPU acceleration for remote access.
As a user with a homelab who deals with a bunch of used enterprise VDi capable Gear & some gaming shenanigans, I am frustrated by subscription models and I often have to tackle that even as a simple hobbyist. NOT having a subscription is VERY appealing. The real special sauce is... CAN we use this hardware in proxmox? Perhaps eventually.
Current set up is dual 2699v3 + 5x tesla P4s running 1 to 1 passthrough on 5x Win10 VMs in Proxmox
I'm running GPUs with passthrough on my current Proxmox cluster but not for VDI, though I've been interested in trying it. Something like this really piques my interest!
I have VMs with GPUs I can use for gaming and video editing. The problem is all the remote access programs I have tried suck and have too much latency. Also wish I could easily virtualize GPU resources. Granted I am still running esxi...
looking forward to proxmox & flex gpu!
This seems really cool, we do small clusters for clients with remote desktop and i wonder if this may increase performance for the end user. Would be interesting to test once the software is there for proxmox :D
So not for enterprise but I do use a VM for my personal machine. It is nicer because in many ways I can access desktop hardware on any cheap laptop. 15 year old macbook air running linux can be a client and run fine. I personally use truenas scale instead of proxmox because the server I am doing this on is primarily a storage server but it works alright.
Over the last year we have migrated a lot of servers to SAAS or into Azure which dropped out vmware count. By our renewal cost though you wouldn't know it. Our VMware hosts are coming up on age and will either be replaced by something else(prob hyper-v) or we might just move the remaining stuff to the cloud. We renewed VMware for this year but I don't see us renewing past next year.
Yet another video showcasing the usefulness of a full height GPU in a server, but I'm stuck with my 2U with only vertical slots (so half-height).
I wish I could find a generic adapter to convert some slots to horizontal full-height but that magic seems relegated to only servers with proprietary motherboards.
Flex140?
The thing is that as soon as they reach better market penetration, execs will force the introduction of licensing for this
Is there Bifurcation on that slot? That would be nice to rig something up so the GPU can still be able to use another 8 lines for something else. Only thing that will stop that is the 1u format of the case. Shame it looks like a very good servers.
Question how well could this work as a media or streaming server and would this technologies go downstream towards the more prosumer side of things? Intel has been doing good work with their video encoding and Im thinking of building a pc for that. It seems right now the use case for this is remote desktops which is great for enterprise but loved to see how this technology works for more local uses.
RDP hw accel is essentially just the same thing as live media transcoding, so it'll work as good if not better because it's not heavily time dependent. I'd imagine you could even do both at once through a little bit of shenanigans.
Ultimately you just need the drivers.
Thanks for the learning, off to see if I can get one for my DL380 G10.
Yeah... We've been VERY CAUTIOUS/concerned about the broadcom buyout. RHEL Openshift is on our roadmap so we're slowly flexing up to shift platforms away from vmware to something else. (we just signed the renewal in December so we're locked for another 2 years safely at the moment).
I do VDI for Personal Use with 2 Tesla P4s used to have consumer Cards, but those are just the sweet spot for light compute and gaming.
I use linux on my laptop but wanna Game sometimes on the go thus I have a couple Windows VMs for me and friends and a Plex using the 2 P4s
This is a solid.... Flex...
I was looking for this comment! After seeing your other videos on VDI, I immediately thought of you! Really excited to see what/if you do anything with this!
so it's vGPU but, not vmware and with intel GPUs?
Stay tuned for part 2, yep
Well one thing that came out of the broadcom deal, it is finally forcing my organization to get rid of the asinine (imo) deployment method we had. We had a bunch of small servers that each had VMware on them, with a single VM each. As the DevOps guy, they didn't let me touch anything even at the OS level let alone hypervisor. But now they're just going for a proper coreos pxie boot setup and both the performance is better and my program isn't going to have to pay the infrastructure team for the convenience of their asinine decisions.
I bought a tesla m10 to try this out myself a while ago, I had no idea there was licensing involved, hopefully that's nothing I need to pay for to use that card, or I'll resell it, and I planned on using an a770 in my machine to try av1 streaming anyway
What do you think of using this in a production studio to have a VM for each Remote editor?
Have an office NAS. Each VM 10GbE to a NAS.
Remote editors run editing software on VM and remote edit through their remote connection to the VM.
Would this work a lot better than using Parsec for remote editing?
vSphere Essentials Plus Kits, got 3 of em, because... reasons. Now that we cant upgrade anymore I wonder if something like this will still work?
I have a A770 still in his sealed box waiting for your secret project 🤩
Proxmox VDI YES PLS !😍
Work with a company that pays a lot for Nvidia license on hardware acceleration for remote desktops for engineers. The whole VMWare/Nvidia licensing is pushing for a change and this is something to consider. This would be great for PCI DSS compliant environments as well they are best served by remote desktop system.
Another VmWare victim here. Will watch while Broadcom is cutting VmWare apart TCM-style. I'm not convinced they won't even let the virtualisation products die a slow and agonizing death. Also looking forward to the class actions once Broadcom inevitably flips all the vSphere-customers the bird ;)
There are going to be fireworks that's for sure. My distributor is a major licensing disty in the UK, and broadcom just killed their partner status. Just out of nowhere.
I use a pair of Tesla P4s within my Proxmox server to provide GPU accelerated gaming VMs to some of my friends. I would love to have a better solution, with proper cooling, and without unnecessary licensing... The whole vgpu drivers thing paired with that they don't even have their own VDI solution, and the streaming part has to be figured out manually is just a joke. I am gonna buy an Intel GPU as soon as it has at least Pascal level of performance and supports SR-IOV without licensing/subscription/whatever... Especially since Nvidia drops Pascal series Tesla card support from GRID drivers starting from version 17...
Removing that Subscription is a Blessing . many times i cannot keep the services paid and need to be able to still acccess services , its only at certan times as budget allows but still down time
Would it be possible to use Intel Arc cards for accelerated VDI as well? Or is limited to the enterprise cards?
So, I am going to be that guy. Faculty is an ability, facility is a place.
I understand the security need for protecting IP, but even with HW acceleration, you are requiring users to remote in to a clean environment with locked down permissions just to then SSH (through putty....) to a head node. Slows the pace of prototyping dramatically
You can do the same thing with SYCL with intel OneAPI
I want one of these so badly.
We call them blade servers at our NOC ☝️🤓
So as a sysadmin that has only ever deployed classic stile RDS solutions on Hyper-V and VMWare.
Does this mean anything? Can I stick in an intel flex GPU in ESXI or Hyper-V hosts and utilize them in RDSH?
Not enterprise, but at home I have a EPYC CPU running esxi8 and have a dedicated "Steam Gaming VM" that passes through a 4060 that I stream to my handheld devices. It works perfectly but would love to be able to share it with my other VM's on the hypervisor. Unfortunately PCI only works with one VM only but I'd be open to installing a second if VMware allows for it to share its resources to any VM that requests it.
I'd love to know the noise level of this Pizza Server. Great video btw
OR some of us use HYPER-V with HW acc... without the nonsense of NVIDIA and VMWARE, just saying
This isn't just about VMWare though, it's specifically about SR-IOV which is a hardware feature that allows much better graphics acceleration by giving each VM its own chunk of real GPU, rather than translating and moving rendering calls in and out of the guest with a ton of extra overhead to prevent VM escape
This is a super cool stuff! Please don't just focus on Windows clients though, Linux should get some validation love, too; Also perfectly on time, as I was thinking about building some VDI infra as our company is growing. I think I still have 6-8 months until I need this, but it would be great to use the Intel Flex and Proxmox combo.
since its made for encoding/decoding i wonder if the Flex would work with Plex hardware encoding
It can serve a whole apartment with Multimedia!
28 1080p streams!
I have been looking into ARC, Ive also heard the built in Intel iGPUs on the latest CPUs work pretty well.@ChuckNorris-lf6vo
Really wish someone would make a card that only did media acceleration. I have a Quadro P2000 for Plex and the drivers under Ubuntu just mysteriously crap out and disappear every few months. I spend a whole day trying to get it working again. Was looking at the ARC GPUs and the better ones need PCIe power. Well, my SuperMicro 3U chassis doesn't have any spare power to run a GPU. I don't need all the horsepower a GPU has to offer just to be wasted for media decoding and encoding.
@@timramich for PCIE I was going to use molex to PCIE adapters.
Also I don't think we really need power we need encode/decode units and VRAM
Would this work as a "Workstation" server for a small office/home? Could I install one of these at the house and then everybody in the house simply have their own remote desktop instance?
Yes, but it would be bad value for money for that purpose if the specific appliance being discussed were used, 5000 dollars buys a lot of clients and the Flex 170 isn't actually that powerful in terms of raw rasterisation
@@bosstowndynamics5488 I was more referring to the GPU in the video. I'd actually like to re-purpose my Threadripper Pro, which I'm currently using (and entirely underutilizing) as my workstation. I'd like to build some kind of machine that can host about 5-6 remote desktop instances that have virtually zero video latency (on the LAN).
@@jeremyrangel8138 Yes but again, the expense. The Flex 170 is an extremely expensive GPU (a quick check of eBay found a single example listed for nearly 2800 Euro, which is more than half of the cost of this entire system). And remember, the 170 is really just an Arc a770 with the GPU equivalent of virtualisation extensions enabled, so it's not even a particularly powerful GPU. 5-6 is right on the edge but I'd be willing to bet that you'd be slightly better off in terms of price to performance by running 6 low end dGPUs and dealing with the slightly more complex wiring/a big case with some Oculink adaptors/PCIe breakouts (6 GTX 1060s could be had for a lot cheaper and would easily run circles around 1/6th of a Flex 170 each even with only a few lanes of PCIe to play with). You could pay for convenience here but this specific server is probably going to be close to the cheapest way to get a Flex 170 build up and running.
@@jeremyrangel8138 Not sure if the other comment is showing but with a TR Pro on hand already your easiest route is to find a decent example of a single slot GPU and slap 6 of them in, you've got plenty of PCIe to spare and 6 cheap GPUs will both cost less than, and run a lot faster per user than, a Flex 170 sliced up 6 ways. The 170 only really makes sense if you want to provide a little bit of GPU to each of 10+ clients
@@bosstowndynamics5488 Are you talking about connecting each user's monitor/peripherals via cable? If so, that's not really feasible. The farthest user from the rack would be about a 100ft cable run....