Gaming on my NVIDIA Tesla GPUs - Part 1 - NVIDIA Maxwell
HTML-код
- Опубликовано: 24 май 2024
- Grab yourself an Insulated Coffee Tumbler at craftcomputing.store
Support me on Patreon and get access to my exclusive Discord server. Chat with myself and the other hosts on Talking Heads all week long.
/ craftcomputing
Nearly four years ago, I started looking at VDI (Virtual Desktop Infrastructure) through Nvidia's GRID tools. It's a system that allows your to bifurcate a GPU, essentially running multiple 3D Accelerated Desktops from a single graphics card. While most official uses revolve around Desktop and Office environments, I've always wanted to use it to play games. And thus, the Cloud Gaming Server Project was born!
As such, I've collected a HUGE number of Nvidia and AMD Enterprise GPUs, and over the next few videos, I'm going to be benchmarking them all. Head to head, which GPU is worth your money, and which ones should you leave on the eWaste pile.
-- vGPU Installation --
Proxmox vGPU Installation Script: github.com/wvthoog/proxmox-vg...
PolloLoco's vGPU Installation Guide: gitlab.com/polloloco/vgpu-pro...
Manual vGPU Install Tutorial: • Proxmox GPU Virtualiza...
-- vGPU Memory and Frame Limiter Unlocks --
(Under 'Define GPU Profiles')
drive.google.com/file/d/1Ok1d...
Links to items below may be affiliate links for which I may be compensated
-- GPU Server --
Asus ESC4000 G3 GPU Server: ebay.us/ngCIDO or ebay.us/hAThwr
Asus ESC4000 G3 Sliding Rails: www.neobits.com/asus_tool_les...
Intel Xeon E5-2697 v4 18-Core: ebay.us/6nlzsD
8x32GB DDR4-2400 REG-ECC Memory: ebay.us/4DNJ00
1.92TB Patriot Burst Elite SATA SSD: amzn.to/4b8Q9jc
-- GPUs --
Nvidia Tesla M40: ebay.us/824gH4
Nvidia Tesla M60: ebay.us/NNS4Ql 
Наука
I’m excited for the P version of this video as I feel that is the best spot right now
It would be awesome to see what a couple of P100 or P40 can do.
9:47 _"I think it's finally time to say goodbye to Kepler, altogether."_
Considering it's been almost 3 years since Kepler support was discontinued, likely Maxwell will be dropped any day now too.
I half expected Jeff to stuff this thing with Pascal or higher.
I just sold off a couple Maxwell based Quadro's before I expect their value to take a dip.
Pascal is in it now.
Volta and Ampere on standby 😉
Was thinking the same while Craft listed the cards off. Not to throw shade, but the Pascals really are the minimum I'd recommend for VDI tasks just as a matter of power and performance. Much like their consumer equivalents they're a quantum leap over their predecessors, so I'm glad he's stepping up to the Pascal cards for the next shoot out.
From my experience as a budding CSE: The P40 is a VDI monster, with 24gb dividing neatly into 2x12gb, 4x6gb, or 8x3gb VMs. The P100 with it's 16gb of HBM2 is still a price/watt/performance monster for Blender or budget AI, It's VDI is faster too, but you waste VRAM in the process. The P4 is in practice a GTX 1070 with a 1080's die, absolutely stompy for a 75w half-height, but I've never tried it and Craft's last video on the P4 wasn't promising.
Pascal support was already dropped in vGPU 17.0 so...
One of the reasons why I only bought three Tesla P4 for my ProxMox servers. Might look into P40 depending on the benchmarks in the next video.
Pascal was 'just' dropped by v17.0. Still supported with the latest drivers from Nvidia (Desktop v535, GRID v16.4) as of 2 months ago.... Not exactly a deal breaker on a 7+ year old GPU.
Aside from gaming, Tesla P40s make for a cheap, though slow, but still useable alternative for hosting LLMs. 24GB VRAM helps.
True, plus they are guide cheap. Least compared to other options.
I have a rig setup like that with twp p40s. The issue is the FP16 performance is atrocious so its very slow to run those models. I've looked at p100s due to that but would have less vram but should be much faster.
Unless your LLM inference framework requires newer CUDA libraries. Nvidia is a horrible POS for making older drivers not work properly with newer cuda libraries.
What makes memory usable in this case is the bit bus. Yeah you can allocate a lot of stuff to that memory but it doesn't help that much if the memory bus is not good.
Really looking Forward to the Pascal results. I have a p40 I haven't been able to test yet but am hoping I can make a low cost VDI server for small and low demand CAD clients.
Happy to see the Tesla vGPU series is back! It has been my fav series in the past. This got me into so many discoveries in the past. It would be nice to see more detailed installation instructions as setup can be challenging
The easiest setup right now is running the vGPU Proxmox installation script here: github.com/wvthoog/proxmox-vgpu-installer
It installs the vGPU unlock, I stalls drivers, and even sets up a virtual license server.
Good experiment setup. Love these proper benchmarks
I've been looking forward to this series! Thanks for the great content, Jeff!
By default, the Frame Rate Limiter (FRL) is enabled for all GPUs (60FPS). The FRL is disabled when the vGPU scheduling behavior is changed from the default best-effort scheduler on GPUs that support alternative vGPU schedulers.
The Tesla M10 is actually stronger than the GRID M40 by 20%-25%.
Thanks for the video! Looking forward to the pascal results. Tesla P4 is a powerhouse of a LP gpu
Good run through. Looking forward to the Pascal tests.
Awesome video! I love the idea of a virtual gaming PC!
On another note, can you do a super quick video breaking down the naming convention for Nvidia/AMD server cards? Specs aren’t usually too hard to find, but broad descriptions of series’s and product lines are extremely rare
Look forward to the P40 test as I have a couple of these in a X99 desktop board with E5-2597 V4. Mostly been playing with AI inference but spinning up some local cloud for older games seems like it could be interesting.
Let me know how it goes. I been thinking of getting some p40. How do you think it would do with local llama3 7b...
@@Zeroduckies P40 does not support FP16 "half precision" so I'd suggest looking instead at P100 16GB that does have hardware support of FP16
16:40 Helldivers 2 is quite CPU heavy. It's not an apples-to-apples comparison: you're probably seeing the more modern 7840U cores pull ahead of the Broadwell cores, despite the Tesla GPUs being faster than 780M in GPU-bound scenarios.
My 9th gen i9 gets really toasty. I’m seriously thinking about converting to water cooling.
I'm actually doing something similar but a bit less technical. I plan on upgrading my powerful gaming PC so it can handle any game I want to play, but I have a pretty small office space and I don't want to play games in a sauna... so I'm going to put my gaming PC in my garage and use steam remote play to display the games off a smaller PC actually in my office. Your method seems to be quite a bit more compact than the 4U server chasis I found for my computer 😅.
Also, during my testing to make sure my setup would work, I found that Steam Link (an older version of Steam remote play that you can install on Raspberry Pis) doesn't like AMD graphics cards. It works just fine with NVIDIA cards (I don't have any Intel cards to test) but AMD cards do not work well😅 I think it's an encoding issue.
Awesome video by the way! Keep up the awesome work!
For starters, love the testing it's good to have an idea where different cards land should anyone dive into this abyss.
Interested in seeing a breakdown of sorts not only on how this compares to building/buying/supporting multiple desktop machines versus a single rack mounted server with VMs. As you mentioned towards the end there, power consumption against performance also plays a big part for some folks so trying to at least mention efficiency would be good to see as well.
Most interested in changes across your setups from 8 VMs dropping down to maybe just 4 VMs or even 2 VMs on the simpler cards. What numbers might you see if you ran say two of the M40s for example with just two VMs running?
But again, thanks for helping me consider something that been in mind for years, just not in patience, could I get the kids gaming with me without having to break the bank...
If you dont mind my asking, What heatsinks did you use in your server? I was following along your build with the ESC 4000 g3, but I am unable to find any heatsinks to use in the system. It's the last step I need before I can boot the server and start on the software side lol
To solve a previous delima about a z420 water cooler, you need to snip or cut the 5 pin plug-ins botton side and just slide it on the 6 pin cpu fan header starting on the top pin. I did it and it actually to my surprise works great. The 140w 6850k i7 I have never gets above 60°C on high end AAA games like Helldiver2 and usually runs in the 30-40°C range, where before it ran up tp 100°C and sometimes would crash while performance would drop off. Amazing all in one water cooler I must say.
I would appreciate it if you added some highlights to your data grids when you call out specific results.
Definitely interested to see the Pascal version, as I happen to have one laying around. Not something I'd want to do personally, but it's fun to know I COULD...
The question is, have you ever tried butter in your coffee in place of cream?
Just hearing a TF2 Soldier Voice line at the start made me like the video
Have you looked at Hyper-V and it’s built in GPU partitioning that also works with consumer cards? I use it in my home server for various VMs running Plex, Frigate and one or two other video decoding/encoding activities. No special mods or drivers required and it works with both Windows and Linux VMs
I'm rocking a Tesla p4 in my main desktop for playing games and vr games (works with quest 2) though I have an older driver so I can bypass the vgpu stuff.
I love my M40 even though I don't use it much anymore. I don't understand why the Pascal stuff hasn't really dropped much in price though. It seems like it's holding on for dear life.
AI homebrew/small bussines machines.
Agreed - AI given they have lots of memory - not fast but you can at least try out CUDA
@@oscarcharliezulu p40 24gb is now AI sweet spot, with good speed and go memory buffer for biger models..
@@gacikpl it is weird because the 24GB M40 isn't much more than the 12GB version. Although the MI25 has jumped in value for reasons that are beyond me. I felt bad about buying it for $65 then watching the price drop to $50. Now it is worth $140... 🤪
@@gacikpl im going to be picking up 4 of them for a localhost model and for a device im developing
Maxwell is not dead yet, it has D3D 12_1 support, and the lattest drivers still support them. U could still run fine any modern game with a GTX 980 Ti.
This server case is exactly what i need to replace my r720xd thats my cloud gaming rig right now. Im running a pair of M40 card setup for passthrough for myself and the step kid.
And this was 100% a gaming video. And i understand and respect that. But while not for this, the k80 is still relevant. If you ever get into playing around with AI, toss a few in there if you have them laying around.
Thanks for the instructions on that by the way, saved me alot of headaches.
Also, i take my daily coffee with a roughly measured 2.5 floz of heavy/whipping cream into about 17 floz of bold coffee. Sometimes i even roast it myself with a cheap air popcorn machine (which you should check out by the way. Makes a terrible mess with the chaff, my wife eants to kill me when i do it indoors, so dont do it indoors)
Love your videos i got this server and some tesla p40s finaly got everything running is there anyway to do 144hz parsec streaming i cant find anything on it other than a dummy plug for consumer cards is there any other option since you cant plug anything into the tesla
Maybe it could be good to also post hashcat benchmark for those cards and other that you are testing
I think it could be nice to include the 780M in all of these bechmarks (or other integrated graphics) so that way there could be a baseline to compare all of these numbers to.
With the M60 would it be possible to use it for Jellyfin and then also a separate VM for Frigate and homeassistant?
I can feel the power draw
way back then in early 2000's 2010's, when I was younger I'd thinker quite a bit with virtual machines, trying new different systems, I flirted quite a lot with linux, and since gaming wasn't there at all yet, I wouldnt just leave windows and distro hop, as fun as it was I couldn't really do too much, it's crazy what you can do with virtualization nowadays!
Is that a Radeon VII in the background on your wall? Still rockin' one in my desktop after all these years.
Noticing that most of these old compute cards are cheap due to new AI services not supporting the older cuda version. Triton server for example supports cuda 6 and newer (going off memory here), and the tesla m40 24gb is cuda 5.2 . Did learn this the hard way too since nvidia docs weren’t clear, so wasted money on a card
Oof. Someone stopped me before I bought a card that couldn't support the version we were running(11), but you're right, they don't make it easy to understand which card supports what.
With these being so cheap I am wondering if just using them for transcoding would be useful.
Just wanted to ask you to test the P4 when you said that it is up next, since my server only supports single-slot cards, thanks
The P4 is definitely in the lineup 😉
Can the CPU low clock be the reason for the low 0,1% framerate?
I think it would be a great idea to do a similar video with these cards in Machine learning, Deep learning, artificial intelligence.
Could some of the stutter be network bandwidth/stack? Is there any image compression from the VM to the client? Is the test repeatable on different client hardware? Sorry for the questions, but it's an interesting experiment that you're doing.
The benchmark/frametime recording was happening locally on the virtual machine. Benchmarking the stream quality would be another metric entirely.
@@CraftComputing Very sensible, thanks for the response
Strange one for you Jeff: I know there's no official drivers, do you know if there's a way to get Maxwell Tesla cards (M40 etc) running on Windows XP and/or Server 2003? I know you can get Titan X Maxwell and the like going with some driver mods, but I've never seen it attempted on the Tesla cards.
Nope. Not heard of anyone attempting to use Maxwell on XP.
@@CraftComputinggtx 960 works out of the box with xp, and simply adding the devixe ids of 970-quadro m6000 24gb gets em working. No idea on the tesla, though it would be pretty awesome especially for splittinf it up for multi vms
great video.
Now I want to see a 16 player classic deathmatch on one server! ;)
You could probably get all 16 onto a pair of tesla P40s, since they scale to 8 3GB vGPUs with the unlock script. Quake 3 Arena and Unreal Tournament both fly on even an 8800 GT, so it should lock-in at ~60 FPS across the board.
Is there anything I can do to calm the beast of fans in the server. I impulse bought one after the first video but I can't handle the noise and nowhere to move it
What about SFF gpus? I’ve got a stack of 2U servers that have gtx750tis that I’m wanting to upgrade
I'm still trying to figure out what to do with the Tesla P4 I impulse bought for $50 a while back.. thought it could be useful for something eventually lol
Any chance you'd try to create a DIY 8xP40 Server. I'm using a 4xP40 Server with an AMD7551p and 512gb of Ram for LLM Inference. Now that Llama 400b is upcoming I am wondering if I still could get some decent performance at 4bit with 8 or even 12 P40ties. Vram is just insanely expensive. Is there a better enterprise GPU than the P40 with the same amount or more Vram and at least the same performance, which does not cost 10x the price of a P40? 😅
How many Patriot burst SSD's you had die on you yet? (I've had a new one last only 2 days!)
Been running these 8 drives for ~18 months now. Zero failures. What are you running them on? What config? RAID?
Would be nice to have some metrics while the game is running like cpu/gpu utilization, ram/vram usage and so on to better identify the bottleneck, maybe the cpu was pinned to 100% and can't keep up or viceversa
I run a dell poweredge r720. Have roughly 128gb ram. Right now 1x tesla p40. I want to get another card. You take these cards and try to split them up into multiple vm's. What about if you created a super vm for gaming on. Take 2 or more cards and pass their performance through to a single vm and give it a ton of system ram and a bulk of the cpu cores. Would that be possible?
Would love it if you also did an ai benchmark on these cards
I virtualized my Pc using hyper-V in one of your previous videos, but I am stuck running at 1080p on the hyper-V windows 11pc i am using via Parsec. How can i increase the host resolution to say 1440p or 4k ? this doesn't seem possible ;(
You should test the tesla T4 and A2, I really want to see those low profile single slot cards
I've previously reviewed the A16, which is the A2 x4. Unfortunately, that one was a loaner, so I can't run it through my tests right now. But my T4 is in the server for the next round.
@@CraftComputing oh I'm quite excited for the next one then, didn't see the a16 video I'm gana have to go back and find that though
Few things to mention:
1. You should probably measure GPU and CPU utilisation, and CPU utilisation per-thread. Then it should be more clear if your low 0.1%s are because of CPU or GPU.
2. CPU is indeed way underpowered. Modern games can easily use 6 cores or more and therefore it is not a good idea of giving a VM less than that. I'm not sure why not to temporary give a VM way overpowered GPU and find sweet spot for the CPU in terms of cores/threads per VM, and then use that as a base.But overall even 3.6 GHz boost clock might not be enough.
3. Another important (for dual-VM tests) thing is PCIe locality. You should always ensure that you allocate VMs to the same CPU that have the PCIe Root Complex where your GPU plugged in. Otherwise you might expereice, let's say, not so great performance scaling because your other VM will need to go via UPI to another socket and you can easily saturate the link.
lol I was actually looking at these on eBay this morning and thinking about this exact question
Wait for the Pascal benchmarks before buying 😉
Hi there, was anyone able to get the vGPU driver with virtual desktops to work with Linux guest VMs?
I suspect the issues with newer games stuttering on these older cards is because the drivers are not tuned to run said games since Nvidia have long-since ended support for cards of this kind of age.
Definitely.
No finds this odd with DX12???? since was supposed to be a low-level A.P.I in which the driver wouldn't be a biggest part of it, yet it's the complete opposite of that????
I have never gotten Sunshine and Moonlight working together properly.
Tried using it to play Halo Infinite over local LAN and it never really worked properly.
It wouldn't take the inputs correctly, which made trying to aim in Halo impossible.
Hi i try my Nvidia Tesla P100 on my HPE DL380 Gen9 Server and folo both type Proxmox vGPU Installation Script & PolloLoco's vGPU Installation Guide. i still cant find my P100 Mapped Device and MDev Type gray out. i try Proxmox 7.4, 8.1 and 8.2 with Kernal 6.5 all fail no see Mapped Device.
What is the input lag? 😟
Craft computing for alcoholics/non drinkers, join me with a cup of coffee or tea! Enjoy your sobriety while also enjoying his content. Much love jeff!
I would love to see some benchmarks for Ai/Stable Diffusion!
Love the idea, I use parsec on my main rig to play games on the go, or to play on my living room TV. Having a server dedicated for game streaming would be pretty fire
Instead of giving each host a dedicated virtual hard drive, could you host your steam library on a NAS and share that with all 8? Are there any issues with that?
The issue is disk I/O if 8 clients are attempting to access the same data.
VHDs on an array mean each host has a unique copy of their own data.
Shared library means only one copy exists, creating a potential bottleneck. I've dealt with that many times in the past when building similar setups.
Hadn’t even thought of Disk I/O. Thanks for the reply - learned so much from your content 😌
I installed my steam library on a FusionIO something pcie 3.0 x8 6.4TB ssd and enabled that new feature that lets you download games from local pc instead of steam servers, it works well with a 10gbit nic
The Nomad test whips my 1080ti down to the movie like 24fps...
The only thing I can't seem to figure out is how to get access too moonlight/sunshine over the net. I can access it locally, but I'd love to let some chums game on my server.
I use ZeroTier for networking
You'll need to setup forward ports in your firewall to access over WAN. Or you can use Parsec, which handles connection via client apps that establish the connection outside of firewall rules.
which one is for fighting, which one is for fun???
I feel the server cpu was really the limiting factor more than anything... 2.3 GHz isn't great 😢
What about RDP?
I am just wondering how P40s would perform.
You'll find out shortly.
Thanks for the video.
Have a main desktop gaming pc running Linux mint pc specs 5800X 32gigs ram 6750xt. TV pc is a 6900HX with 680M 32gigs ddr5 5200 Linux mint os. Would like to stream games from the desttop pc to the tv pc what software would I use to do so. Oh I do not use steam have not purchased a game there for 10 years. Using Heroic games Launcher using my GOG or Epic games.
Probably parsec?
@@gacikpl is Parsec a program?
@@ehiebert1297 yup for remote desktop with 3d and gpu acceleration.
Since your Desktop is running Linux, and you say you don't use Steam, you'll want to take a look at Sunshine which is actually what Jeff uses on his server. Unfortunately Parsec does not have a Linux Host application.
Note that you can 100% create a Steam account for free and you can use Steam's game streaming to stream your non-steam games 100% free, however due to the way Linux Steam / Proton installs games it might be tricky to get that working properly.
VDI will cause issues with certain anti-cheat software. If you plan on playing online multiplayer with many new titles
0:36 DANG IT JEFF!..You flubbed your lines.. "But what is Cloud Gaming?" - Onlive, 2015
I love it......
Re upload ?
Unfortunately Kepler is the most recent Nvidia chipset still supported by Apple.
I work with someone that is getting 2 K80s in a GPU server for fairly cheap but wants to use it for AI/ML I told them it is horrible idea to use a card that old for something that will probably max those cards out all the time using them and aren't going to be any bit efficient compared to something even slightly newer.
At least upgrade to the M60... they're super cheap, and supported by much more modern software (both vGPU/GRID and CUDA).
@@CraftComputing that’s what I told him but won’t listen. He will regret his choice on it.
@@jasonscherer2631 I just abandoned using a k80 because the cuda support level is only 3.5 and most programs like stable diffusion require at least cuda 5.0 or higher. I upgraded to the tesla p100 for $200 off ebay
Helkdivers2 uses 40% of my 9th gen i9 and 80%+ of my 3060 12gig.
The 7840U has a higher multi-core score in Passmark than the Xeon-E5-2697-v4. In single core it's outright humiliation.
Great, now i want to make a sports coat out of tin foil so i can go as full metal jacket for halloween
lol I love it.
17:02 drivers. Drivers drivers drivers.
At first I thought was watching a Doug Demuro video
THIS is the 2015 NVIDIA Tesla M40, which originally sold for $5,700... and today, I'm going to review it.
Jeff knows too much to emulate Doug.
I don't dislike doug, but whenever he reviews older cars, he knocks them for not having cupholders and is mystified by such things as an EQ in a stereo.
saying something happened 10 years ago doesn't phase me anymore. I'm already in the century-scale for mental time (even though i'm only 38 yo). anything less than 20 years seem a bit insignificant to me 🧐
How about CSGO? Is it playable?
It's time to give up on Maxwell.
THis is the kind of video you get when you have somebody who wants server hardware but doesnt have a server job to give it.
the K80 was the biggest waste of money for me, all gpus are double the price in europe and the P40 while being 2.5 times more expensive has like 200% more perofrmance. also the K80 does horrible in 1080p and is pretty laggy and the mehtod of using the K80 for cloud gaming on this channel only allows for 1080p.
short answer no...
Saving the planet from e-waste, by streaming games ..
Bravo!!!
21:31 those P cards are still so $$ because of Ai.
We run the L4 cards for our PoC vdi, they are amazing..... But alas, absolutely nuts on pricing.
I like watching these videos for educational purposes...
however the concept of streaming games is one that I will never ever support.
I mean i like to watch how they setup the servers and all that, but 100% of the time I would rather run the game locally rather than stream it.
Try playing a ranked dota 2 match through a streaming client......... see how well you do...... ;)
Anyways keep up the videos!
Fallout the TV show is a mixed bag. I personally enjoyed most of the characters and like everyone else, I really like the cowboy ghoul. The bad parts are small but glaring. Luckily, they don't outpace the series. I didn't find myself cringing out like I usual do with most of the modern Disney Star Wars or Marvel. Watch it and let it permeate the senses, but if you are absolutely a purist when it comes to Fallout lore, it'll make you happy and sad at the same time.
'Purist' Fallout lore is hilarious, considering how outlandish and wacky some of it is. If anything, it's a bit toned down in the show.
@@CraftComputing the fact that they got the creator to say "i'm not writing anything more on the lore apart from what i've already written, the franchise is in bethesda's possession now so go by their lore" is like thinking star wars could ever be written by disney
Spoilers:
The bit where he was like "oh I don't want my Pepe to pop" was super cringe 😬
Overall thought it was great
@@hotrodhunk7389I guess this what happens when reddit don't teach you how to rizz
@@damasterpiece08I mean, star wars animation is on a roll lately. Last season of clone wars, bad batch, tales. Everything is a banger.
yea veri nasti !!! so go back gtx 500 fermi ..and let htm fix tha sofwer probelsm man if they dodnt som one els weil !!! AMD /INTEL ...ETC
Tesla Features
I always think of 22 Jump Street when he says "I am Jeff"
"modern games" ban you for running in a vm
Only those without support for AntiCheat. There are some that work, there are others that don't.
Beating on a dying horse ... leave it already ... you made like 5 videos on this ...
I made like 20 videos on this. It's something I'm interested in, and want to share every nuance I can. If it's not for you, don't watch, but this is also my most watched series on the channel. Don't assume other people don't care just because you don't.
@@CraftComputing So you advise me to not watch and not comment, because i might be in minority of people why are kinda over this. But you are absolutely right, this might not be for me anymore. As of late your content has gotten stale, boring, repetitive and overall not fun as it used to be. Now since im sure im not only one that feels this way, at least im one that express my feelings openly. I am thankful for most of previous videos and quite a bit of them are amusing and were fun to watch. So for that, thank you.
@@Mihazerodamn who pissee in your serial port this day
there are alot of them... but we do not post them, have many - trying to make things easy for the common man... you - on the other hand - try to make this difficult, if you combine common sense - and a little google everyone will make this work,,,
....wtf????
Fallout the TV show is boring and each episode is full of filler.
What's it matter to you if I enjoy a show you don't?
can you benchmark games with a modern cpu
For the low low price of $2500, I could get a more modern platform with higher clock speeds to run these cards. Until I win the lottery, you'll have to live with the hardware I can afford.
Great video as always. It certainly sparked my interest once more...
I've had a M40 24GB for years that never had any use, but I couldn't bring myself to get rid of it. I saw some of your videos about vGPU and attempted to get the card working but got the following:
NVRM: GPU 0000:03:00.0: RmInitAdapter failed! (0x24:0x72:1417)
NVRM: GPU 0000:03:00.0: rm_init_adapter failed, device minor number 0
That error stopped me and I wasn't able to figure it out, thus the GPU card went back in storage. This video made me pull it out once more, though to no surprise the same error. Searching didn't provide anything fruitful other than the GPU card "may" be defective, but not a definitive answer. By chance have you (@CraftComputing) or anyone see this error?
I've not seen that error. Try installing drivers using this script: github.com/wvthoog/proxmox-vgpu-installer
It also configures Proxmox with IOMMU, and installs the proper drivers. Give it a shot and see if that's your issue.
@@CraftComputing I had been developing a script to streamline the process, guided by your video and several other sites that provide a walkthrough for this task. So, I'm surprised but pleased to find that the work has already been completed and is now accessible.
Despite the excellence of the script, I'm loathe to report that the aforementioned error is likely a hardware related failure. I've determined such after taking the card out and into a windows server (loaded with the correct drivers for it) where sadly it did not appear. Thank you for your quick reply, I'll look into a replacement and hopefully this can help others if they encounter it.
Do a search for "rm_init_adapter failed". On the NVIDIA developer forum, user furrious worked past this error on his 1650, by adding a command line parameter -m=kernel-open. Not sure if that will help you, but if you haven't tried it, it's worth a shot.