[MOVED] Hacking NVidia Cards into their Professional Counterparts

[dgusev]

Yes, it is official now. GTX/RTX is enabled for passthough without modding from R465 beta drivers.

Awesome! Personally I never had code 43 issues, GTX passes through fine and driver doesn't care at all about VM. 

I think we'll definitely be looking at a more open future of consumer graphics cards. Also, for those of you are working remotely with GeForce cards, Nvidia released a patch you can download called "nvidiaopenglrdp.exe" that enables remote desktop OpenGL for CAD users working remotelye. You'll need to create an Nvidia developer account to download it.

Edit: @dgusev my P106 is broken so I cannot try the 3D enable mod similar to how you did with the Tesla K10. It may work... but for some reason the +5V line is going straight to GND. 

Looks like capacitor short circuit

[krutav]

Looks like capacitor short circuit

Probably is... I don't have schematics for this model so I'll never know. Anyway, I got good use out of it for the $16 it cost to buy P106 back in the day. I don't have any other 3D disabled graphics cards to test with, and I don't see any purpose in shelling out $200+ for a Maxwell Tesla because all these consumer GeForce cards are far better in everything but OpenGL. However, the Nvidia Studio driver which supports Pascal series graphics cards does have some OpenGL optimizations I think, and the performance in Blender is pretty good so I don't worry too much about buying a Quadro just for that OpenGL boost.

As far as the P106 goes, a 12v to 5v step-down converter bust and I'm not sure what that did because the surroundings are in perfect condition. I'll need to get GTX 1060 (Zotac preferably) schematics/boardviews to get a better idea of how to go about fixing it, so I'll be on the lookout for some of these files, and I can get back to experimenting!

One big problem: P106 has no display outputs or NVENC, so I don't see any method of actually enabling normal function with it even if I use the @dgusev vBIOS 3D unlock method...
Edit: Only Quadro can create virtual display instances so there's that.

[dgusev]

I recently installed Nvidia OpenGL RDP fix that is not required for Quadro cards... 
So far, before this dumb move from my side, GPU had no issues in virtualized gaming and even CADs. Everything worked as promised.

All kinds of VMs ended up with unfixable Error 43.

Result: bricked vBIOS on GPU #1 (the only that was working hahah). Cannot reflash it using nvflash. Waiting for my CH341 to force flash chip.

[keggerbra]

Hello Everyone! My name is Adam, found this thread randomly and thought I would pop in.

This sounds like an amazing project you guys are working on. Has anyone here considered trying to flash or convert an AMD FirePro S9150 series card? I have a few available and would be willing to dedicate them to the effort if anyone is interested.

I have a professional background creating VDI architecture for hospitals and large scale production environments. These cards are currently part of my home lab and I used them for proof of concept for a technology patent. However I’m in the same boat as you guys, I like to try new things and experiment! While I get to work on some serious hardware, I’m limited in my homelab and the current state of GPU availability, application/OS compatibility, and licensing is abysmal. I have various projects on hold due to limited options when it comes to options.

I would love to see the capability of these cards running as mxgpu in VMware ESXi, KVM, Xen and if I can help contribute to make it a possibility, please feel free to contact me or reply. Thanks!!


https://www.amd.com/en/support/professional-graphics/firepro/firepro-s-series/firepro-s9150



Some more pictures of my cards
https://imgur.com/a/sl6iIDM

[krutav]

I would love to see the capability of these cards running as mxgpu in VMware ESXi, KVM, Xen and if I can help contribute to make it a possibility, please feel free to contact me or reply. Thanks!!

Hello Adam,

Unfortunately, the FirePro S9150 is a GCN 2.0 GPU marketed towards VDI, but it does not support SR-IOV. The MxGPU feature requires SR-IOV functionality, which was implemented in the next generation. This new generation is GCN 3.0, more specifically the "Tonga" AMD chip in the FirePro S7150s supports SR-IOV and MxGPU as a result. So you probably aren't going to be able to get MxGPU on S9150 sadly.

There are other options though like VMware vSGA shared graphics, but that is old. New in Microsoft's HyperV is GPU-P which is the successor to RemoteFX and works on most cards, including the FirePro S9150. It works really well, but it's only an option worth considering if your environment is primarily Windows.

[gotofbi]

Based on my testing, the internal timer to license requirement starts at 20 minutes of guest VM Nvidia driver running. If you can mask that, you get the features forever. But that is harder to do.

The guest's driver? On a Windows guest? Then here is a crazy idea: Attach CheatEngine to the driver, enable speedhack, speed = 0, done. I have no idea if that could work, it will probably cause other problem though.

Quick tip for anyone whos fighting on the license.

1. Forget about FlexNet. Its not that easy to crack after all

2. When Guest(Windows) enters limited performance mode (after 20min), try to disable the vGPU device in Device Manager and start again.
This will reset 20min timer and gives you big hint that where the 20min gets calculated.
I tested this on vGPU 10.2 driver so not sure about latest vGPU 11 though

[gotofbi]

This sounds like an amazing project you guys are working on. Has anyone here considered trying to flash or convert an AMD FirePro S9150 series card? I have a few available and would be willing to dedicate them to the effort if anyone is interested.

I dont think you can make S9150 SR-IOV compatible.

After all, S7150 has special bios for SR-IOV that other card does not have.

My suggestion will be get W7100, put bigger BIOS chip, flash S7150 bios, override PCI-ID in guest OS to allow AMD gpu driver to be installed.

Detailed info can be found here:
https://forum.level1techs.com/t/how-to-sr-iov-mod-the-w7100-gpu/164186

[krutav]

Quick tip for anyone whos fighting on the license.

1. Forget about FlexNet. Its not that easy to crack after all

So me and some other folks running the vGPU_Unlock Discord server and Wiki have figured out how to bypass the license entirely with "x-pci" QEMU argument. I won't go into details for obvious reasons, but this completely throws away license requirements. Of course, if you want to use vGPU I highly recommend paying Nvidia for it, but yes there is a workaround. I may have posted the Discord link here, though if it isn't there, send me a message.

My suggestion will be get W7100, put bigger BIOS chip, flash S7150 bios, override PCI-ID in guest OS to allow AMD gpu driver to be installed.

If you have a bunch of S9150s lying around, what's the point of spending more money? The MxGPU technology on S7150 is not very advanced and Nvidia vGPU is better than it in many ways. My recommendation for Adam is to use them in standard passthrough mode, that way the cards don't go to waste. With 8 of those GPUs, I don't think GPU mediation is necessary anyways. These S9150s aren't very new, nor that great, but they can be used in a remote gaming or workstation setup I suppose.

Perhaps the best use of those cards is Folding At Home, where you can give unused compute power to help with scientific calculations that can benefit society.

[BLEACH]

Has anyone attempted this on the new 30 series cards ive found the resistors by the bios chip ive unlocked the mismatch id by changing it in the bios but keep getting caught by the certification on nv flash im trying to change device id from 2216 to 2206 with the resistors ?

[BLEACH]

im going to attempt this on a newer 3080 but i noticed in nvflash the eprom always keeps the same id number and ive noticed nividia marks them with different paint so i need to know if anyone has a 3080 450-480 watt model 2206 bios and if they can get me the number nividia use on that exact vbios chip in it i will be replacing it as a second bios replacing the switch and trying different resistors to see if i cange 2216 to 2206 thank you in advance !

[Rettep]

Can you do that with the mining P106-100 without having the limitations of the old drivers?

[Rettep]

Can you do that with the mining P106-100 to use it as a normal 1060?

[Dave Boechler]

Have a GTX 690 which resistors have been changed and want to set it back to stock anyone have a picture of the resistor locations and values for a stock GTX 690?

[Xii-Nyth]

Any luck?

[Rickochet]

@Rettep
Any luck? or still going with 422.15 ?
Either this: Directx12 does not crash, but NVDIA control panel does not open: https://github.com/iiiyx/p106-100-p104-100
Or that:   Directx12 games crash,  but NVDIA control panel  opens:  ( change that MShybrid registry key) :   https://youtu.be/dkE4aAVUIzU


I would love to turn P106-100 into a 1060 6GB on a firmware level. Can't we just reprogram the bios EEPROM chips?

Also for some reason my Asus P106-100 has 2 display out ports (DP and HDMI) that are used as screwholes and blocked by the IO-plate. I wonder if I can utilize those ports and get somehow displayoutput from them? ( If the video driver amplifier circuitry is not missing?)

   

[gordan]

Wow, this forum thread has been quiet for a while.

Questions:
1) Has anyone figured out how to resistor-mod PCI IDs on the Pascal and Turing GPUs (10xx and 20xx series)? I am guessing the BIOS ID bit strap modding like in the <= Fermi days hasn't started working again.

2) Has anyone had any significant success with flashing Quadro BIOS, e.g. RTX 6000 / RTX 8000 BIOS onto a 2080 Ti or Quadro P6000 on 1080 Ti?

Context:
I think there may be a new call for Nvidia modding. Specifically, Pascal (10xx) and Turing (20xx) cards are dirt cheap, and the AI craze is pushing the need for lots of VRAM more than lots of compute.
2080 Ti cards have been modded to 22GB successfully and 44GB unsuccessfully. The latter boot, but the driver refuses to touch them.
But if we compare reference designs, RTX 8000 is similar to 2080 / 2080Ti and RTX 6000, but RTX 8000 has memory chips on both sides of the PCB populated for a cool 48GB of VRAM.
We know that 2080 / 2080 Ti has part of the GPU cut out (apart from a few rare notable exceptions), so it can only use 11/12 memory chips.
But the 44GB mod seems to be constrained by software rather than hardware. This shouldn't be surprising given the dodgy techniques Nvidia previously used to block PCI passthrough on GeForce cards.

I'm wondering of a 44GB modified 2080 Ti could be made to work with the Quadro vBIOS and driver and thus use the full 44GB of VRAM (assuming the > 24GB restriction is enforced in the vBIOS and/or driver).

This would give a lot of people a cheap and effective way to run larger models on their relatively cheaply modified GeForce cards instead of having to spend multiples more on the likes of RTX 8000.

[Coldblackice]

I think there may be a new call for Nvidia modding. Specifically, Pascal (10xx) and Turing (20xx) cards are dirt cheap, and the AI craze is pushing the need for lots of VRAM more than lots of compute.
2080 Ti cards have been modded to 22GB successfully and 44GB unsuccessfully. The latter boot, but the driver refuses to touch them.
...
This would give a lot of people a cheap and effective way to run larger models on their relatively cheaply modified GeForce cards instead of having to spend multiples more on the likes of RTX 8000.

Interesting, followed. The "driver refusing to accept them", is this due to some sort of hash match/verification? I wonder if this would be possible to patch out at the software level.

[gordan]

The "driver refusing to accept them", is this due to some sort of hash match/verification? I wonder if this would be possible to patch out at the software level.

Unlikely - driver signing prevents any tampering with binaries. Same with BIOS signing. But there seem to be ways of force-flashing Quadro BIOS onto GeForce cards of recent generations. Back when I was doing this sort of thing in the Fermi and Kepler days, the only artificial obstacles in the way were in the drivers whitelisting PCI IDs.