Tesla P4 - 8GB / US$ 100 or even lower

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

zunder1990

Active Member
Nov 15, 2012
226
78
28
Thats not correct. Thats completely unrelated to proxmox. You are required to license every guest system. You need the host drivers as well. There is no way around execpt the ways I mentioned.

Getting the drivers isnt an issue.
If you talking about vgpu under promox you DONT need a license for anything. There are a bunch of work arounds PolloLoco / NVIDIA vGPU Guide · GitLab
 

FlorianZ

Well-Known Member
Dec 10, 2019
201
258
63
I think we are all saying the same thing here:
  • You're legally required to have a GRID license, even for Proxmox.
  • You don't technically need a license to make it work. There are workarounds available.
  • You need the GRID drivers in either case. They are locked behind the licensing portal, but are still easy to acquire with a trial account or if you're comfortable with downloading them elsewhere.
 

gizzard

Member
Nov 21, 2020
22
27
13
As an update, I purchased one of these cards for ~$80 all in. Arrived in pristine condition about three weeks later. Signed up for the GRID host drivers which come with Guest drivers too. My first inclination was to put it in my desktop and use it in KVM, but I couldn't get the mdevs to appear. Admittedly this was Arch on a 6.6 kernel, I accidentally installed 13.x drivers at first instead of the newer 16.x one, and I didn't use the vgpu-unlock script at all. I then tried it an Elitedesk G2 800 SFF, but the PSU couldn't power the card and a PCIe SSD at the same time (the card also barely fit in the chassis and there would be no easy way to cool it).

So I moved it to my NAS (PVE 7.x, X10SLM-f, E3-1230v3) and it works. I followed the PolloLoco guide and passed an Q graphics profile with 8 gb graphics memory to a Windows 11 VM. It appears as a "Nvidia GRID P4-8Q" in Device Manager. With no special cooling, it idles at ~45 deg C. I installed Sunshine and it jumps to around 80-85 deg C when streaming at 1080p @ 60 Hz. I am having issues with stuttering and dropouts but I think there's more graphics configuration in Win11 to do and the client is connected via 5ghz wifi.

For me it was a pain to configure the displays in Proxmox because the console uses SPICE but Sunshine uses a secondary display that I could only access via RDP, but you cannot change resolutions or other display settings via RDP. Every time I made a change via the console, the "Keep these settings?" box would be inaccessible since it was on the second display. Had to do a little dance to figure it out.

There are instructions on how to set up a licensing server. I am by no means an expert but basically the licensing server issues a token that you provide to the VM. The driver then connects to the server and becomes authenticated for ~90 days.

I was going to use a vGPU for Frigate motion detection, but I am getting ~20 msec inference time at 10 fps just using the Openvino detector on a G4560t (2c4t, 2929 passmark, HD610 gpu). I don't have many cameras so I guess I'll just use the P4 for a gaming server...
 
  • Like
Reactions: gb00s

dontwanna

Member
Dec 22, 2016
93
20
8
With no special cooling, it idles at ~45 deg C. I installed Sunshine and it jumps to around 80-85 deg C when streaming at 1080p @ 60 Hz. I am having issues with stuttering and dropouts but I think there's more graphics configuration in Win11 to do and the client is connected via 5ghz wifi.
The best cooling performance I've managed to get so far was by removing the front plate from the card, and pointing a large fan at the (now exposed) heatsink. Not always an option, since the card with 25mm fan(s) then takes up ~ 3 slots instead of just 1, and might be tricky to keep the fan(s) in place, but temperature- and noise-wise it works great - I went from 88 C (1 fan blowing through the heatsink with the front plate on) to 58 C (1 * 120 mm fan pointed directly at the exposed heatsink fins, with the front plate removed). That's with ~ 60-70 W load reported with nvidia-smi.
 
  • Like
Reactions: gizzard and Marsh

zachj

Active Member
Apr 17, 2019
172
115
43
As an update, I purchased one of these cards for ~$80 all in. Arrived in pristine condition about three weeks later. Signed up for the GRID host drivers which come with Guest drivers too. My first inclination was to put it in my desktop and use it in KVM, but I couldn't get the mdevs to appear. Admittedly this was Arch on a 6.6 kernel, I accidentally installed 13.x drivers at first instead of the newer 16.x one, and I didn't use the vgpu-unlock script at all. I then tried it an Elitedesk G2 800 SFF, but the PSU couldn't power the card and a PCIe SSD at the same time (the card also barely fit in the chassis and there would be no easy way to cool it).

So I moved it to my NAS (PVE 7.x, X10SLM-f, E3-1230v3) and it works. I followed the PolloLoco guide and passed an Q graphics profile with 8 gb graphics memory to a Windows 11 VM. It appears as a "Nvidia GRID P4-8Q" in Device Manager. With no special cooling, it idles at ~45 deg C. I installed Sunshine and it jumps to around 80-85 deg C when streaming at 1080p @ 60 Hz. I am having issues with stuttering and dropouts but I think there's more graphics configuration in Win11 to do and the client is connected via 5ghz wifi.

For me it was a pain to configure the displays in Proxmox because the console uses SPICE but Sunshine uses a secondary display that I could only access via RDP, but you cannot change resolutions or other display settings via RDP. Every time I made a change via the console, the "Keep these settings?" box would be inaccessible since it was on the second display. Had to do a little dance to figure it out.

There are instructions on how to set up a licensing server. I am by no means an expert but basically the licensing server issues a token that you provide to the VM. The driver then connects to the server and becomes authenticated for ~90 days.

I was going to use a vGPU for Frigate motion detection, but I am getting ~20 msec inference time at 10 fps just using the Openvino detector on a G4560t (2c4t, 2929 passmark, HD610 gpu). I don't have many cameras so I guess I'll just use the P4 for a gaming server...
Does the P4–as an honest grid product—need the pollo loco hacked drivers? I’d think it doesn’t…or do you have to use the hacked drivers if you want to spoof the PCI device ID in the windows VM?
 

dontwanna

Member
Dec 22, 2016
93
20
8
Does the P4–as an honest grid product—need the pollo loco hacked drivers?
No, it doesn't require any patching, the official nvidia grid drivers work. That polloloco's guide pretty much covers it all, except how to acquire the drivers (other than the trial workaround, which is tedious, especially if you don't have a "business" e-mail). There's a github page where one can download all the grid drivers freely, not gonna link it directly here since it's probably against STH rules and there's always a chance that those drivers were somehow "tampered with" (unlikely, but still possible), can be found with keywords like "vgpu", "archive" etc.
 

zachj

Active Member
Apr 17, 2019
172
115
43
No, it doesn't require any patching, the official nvidia grid drivers work. That polloloco's guide pretty much covers it all, except how to acquire the drivers (other than the trial workaround, which is tedious, especially if you don't have a "business" e-mail). There's a github page where one can download all the grid drivers freely, not gonna link it directly here since it's probably against STH rules and there's always a chance that those drivers were somehow "tampered with" (unlikely, but still possible), can be found with keywords like "vgpu", "archive" etc.
I’ve looked and I don’t see any detailed tips on how to get the demo drivers…it seems obvious that I’d need a custom email domain but what are people doing to renew every 90 days? Are they just creating a second (and third and fourth and fifth) email address at their custom domain? Or registering for new domains every 3 months? Obviously I know how to do those things but I’m wondering if people have stumbled upon a less complicated way…

im aware of the fake DLS option and the windows registry edit to reset the 24 hour grace period, and I’m also aware I can spoof the pci device Id in KVM so I can use standard consumer drivers in the guest…my problem is VMware won’t let you spoof the device ID and the registry key hack is obnoxious (requires a reconnect to the vm every day) and I don’t love the fake DLS option.

I know beggars can’t be choosers and it’s probably silly for me to be opposed to all three perfectly cromulent options, but alas here I am asking if renewing trial licenses every 90 days is easier :)
 

dontwanna

Member
Dec 22, 2016
93
20
8
it seems obvious that I’d need a custom email domain but what are people doing to renew every 90 days? Are they just creating a second (and third and fourth and fifth) email address at their custom domain?
Yeah, that's the same questions I've been asking myself. None of that made any sense to me. Register a trial account for 90 days, download the current drivers, and then, if you want new drivers later, you'd need another trial account, and so forth. Seems like way too much trouble, so I just gave up on that idea entirely and simply download the drivers from github. Sure, there's a risk they're not "legit" somehow, but that github page been there for months, and I guess someone would've noticed something fishy going on there by now if there was anything wrong with it (all it takes is a simple checksum comparison between the drivers downloaded via trial account from nvidia and the ones posted on github). Other than that, I can't think of any less complicated ways to get current drivers besides registering new trial accounts with new e-mail addresses, and maybe IP addresses etc every 90+ days.

and I don’t love the fake DLS option.
I'm just running a docker in a dedicated VM with FastAPI-DLS, the same kind of compromise as downloading drivers from github, I guess. I don't think it can do much harm even if it wanted to, it doesn't need internet access, so that VM can be locked down completely and only the port required for license renewal can be left open.
 

zachj

Active Member
Apr 17, 2019
172
115
43
Yeah, that's the same questions I've been asking myself. None of that made any sense to me. Register a trial account for 90 days, download the current drivers, and then, if you want new drivers later, you'd need another trial account, and so forth. Seems like way too much trouble, so I just gave up on that idea entirely and simply download the drivers from github. Sure, there's a risk they're not "legit" somehow, but that github page been there for months, and I guess someone would've noticed something fishy going on there by now if there was anything wrong with it (all it takes is a simple checksum comparison between the drivers downloaded via trial account from nvidia and the ones posted on github). Other than that, I can't think of any less complicated ways to get current drivers besides registering new trial accounts with new e-mail addresses, and maybe IP addresses etc every 90+ days.


I'm just running a docker in a dedicated VM with FastAPI-DLS, the same kind of compromise as downloading drivers from github, I guess. I don't think it can do much harm even if it wanted to, it doesn't need internet access, so that VM can be locked down completely and only the port required for license renewal can be left open.
I read most of the fastapidls code and I don’t see anything fishy in it; like there was no obvious attempt to phone home. Most of my heartburn about it is I expect the nvidia guest driver probably *is* phoning home about its license and it wouldn’t be too hard to connect the dots about where a license came from.
 

NachoCDN

Active Member
Apr 18, 2016
123
106
43
54
has anyone looked at this card for running LLM's? (Large Language Models). i can't seem to find any information related to accelerating AI workloads and this card.
 

DaSaint

Active Member
Oct 3, 2015
283
80
28
Colorado
I have mine working with TrueNas Scale, got them to support (works out of box) it so i could pass it to PLEX via Scale Applications (GPU Transcoding) i added a 3d printed fan shroud and mini fan to it to provide some active cooling and it works great (Grant you im not doing 4K) but does a lot of the heavy lift vs the CPU and it works in K8S in TrueNas Scale
 

nk215

Active Member
Oct 6, 2015
413
145
43
50
I'm just running a docker in a dedicated VM with FastAPI-DLS, the same kind of compromise as downloading drivers from github, I guess. I don't think it can do much harm even if it wanted to, it doesn't need internet access, so that VM can be locked down completely and only the port required for license renewal can be left open.
Did you do that under ESXi?
 

gizzard

Member
Nov 21, 2020
22
27
13
These cards are now down to around $70 before fees / taxes.

The best cooling performance I've managed to get so far was by removing the front plate from the card, and pointing a large fan at the (now exposed) heatsink. Not always an option, since the card with 25mm fan(s) then takes up ~ 3 slots instead of just 1, and might be tricky to keep the fan(s) in place, but temperature- and noise-wise it works great - I went from 88 C (1 fan blowing through the heatsink with the front plate on) to 58 C (1 * 120 mm fan pointed directly at the exposed heatsink fins, with the front plate removed). That's with ~ 60-70 W load reported with nvidia-smi.
This is a good idea! I got this 3D printed plastic bracket and a Delta 50x50x10 blower style fan. It is not quiet, but now nvidia-smi shows the card idles at 30-32 C and gets to around 60-64 C with 45-50 W of load. I played a bit of Elder's Ring on maximum settings at 1080p / 60 fps. There is rare stuttering and I am still messing around with Moonlight configuration.
 

zachj

Active Member
Apr 17, 2019
172
115
43
These cards are now down to around $70 before fees / taxes.



This is a good idea! I got this 3D printed plastic bracket and a Delta 50x50x10 blower style fan. It is not quiet, but now nvidia-smi shows the card idles at 30-32 C and gets to around 60-64 C with 45-50 W of load. I played a bit of Elder's Ring on maximum settings at 1080p / 60 fps. There is rare stuttering and I am still messing around with Moonlight configuration.
it’s not quiet? Dang! I was about to buy 2…

is your fan on a fan controller or is it running at maximum speed 100% of the time?
 

gizzard

Member
Nov 21, 2020
22
27
13
it’s not quiet? Dang! I was about to buy 2…

is your fan on a fan controller or is it running at maximum speed 100% of the time?
The fan is 3 pin and connected directly to 4-pin FAN3 port on the motherboard (X10SLM-F). IPMIView shows it running at 5700 rpm which I am thinking is 100% speed. I set the fan profile to "Optimal" but it did not make a difference.

For context, this system is in a Fractal Node 804 and I cannot really hear the three 120 mm case fans that are also 3-pin. I cannot remember whether these case fans are also connected to the board or to a molex adapter. I think the sound from this blower fan projects directly out of the case through the GPU heatsink so the case cannot really suppress the noise.

I might pick up this 4 pin PWM fan instead.
 

zachj

Active Member
Apr 17, 2019
172
115
43
The fan is 3 pin and connected directly to 4-pin FAN3 port on the motherboard (X10SLM-F). IPMIView shows it running at 5700 rpm which I am thinking is 100% speed. I set the fan profile to "Optimal" but it did not make a difference.

For context, this system is in a Fractal Node 804 and I cannot really hear the three 120 mm case fans that are also 3-pin. I cannot remember whether these case fans are also connected to the board or to a molex adapter. I think the sound from this blower fan projects directly out of the case through the GPU heatsink so the case cannot really suppress the noise.

I might pick up this 4 pin PWM fan instead.
Ok that makes more sense.

I took an old nvidia k2000 apart and taped the blower to my T4 and it works and is quiet unless I really ramp up the fan speed. Problem with what I’ve got now is temps get into the 80s even while watching a video.
 

Markess

Well-Known Member
May 19, 2018
1,183
803
113
Northern California
The fan is 3 pin and connected directly to 4-pin FAN3 port on the motherboard (X10SLM-F). IPMIView shows it running at 5700 rpm which I am thinking is 100% speed. I set the fan profile to "Optimal" but it did not make a difference.

For context, this system is in a Fractal Node 804 and I cannot really hear the three 120 mm case fans that are also 3-pin. I cannot remember whether these case fans are also connected to the board or to a molex adapter. I think the sound from this blower fan projects directly out of the case through the GPU heatsink so the case cannot really suppress the noise.

I might pick up this 4 pin PWM fan instead.
I think that any time you have one or more 4 pin fans on those X10 series boards, any connected 3 pin fans tend to run full speed. Do you maybe have a PWM fan on the CPU?
 

nk215

Active Member
Oct 6, 2015
413
145
43
50
No, I'm running Proxmox. Nothing unique about that VM with licensing server though - it's just a clean Ubuntu VM with FastAPI-DLS, so should work the same in ESXi.
I got it working. Thank you.

Here's my observation. At the end of the day, I like the approach that Promox did, with vGPU unlock, much better. In ESXi (and the license server), I have to split the GPU into equal (already defined) chunks. I can't have a VM with 2 Gig GPU memory and another with 4 Gig GPU memory.

In Promox, I can do that, have one VM with 2.5G, the other with 3G, another one with 4G etc w/o any issue.
 
  • Like
Reactions: NachoCDN and gb00s