[MOVED] Hacking NVidia Cards into their Professional Counterparts

[colokarl]

Hi Guys,

got my K10 in the mail today.

As per post (attached pic in the following post):

https://www.eevblog.com/forum/general-computing/hacking-nvidia-cards-into-their-professional-counterparts/msg2899670/#msg2899670

20 k resistors have to get removed and 40k resistors added.

But this conflicts with the following table as per post:

https://www.eevblog.com/forum/general-computing/hacking-nvidia-cards-into-their-professional-counterparts/msg2832956/#msg2832956

Can anyone confirm, the way of first link with pic is the correct one, as it seems to be reading the posts following after that.

Also, i think i need to flash the grid k2 bios after modding the resistors, is this correct and in case it is, which one should i use ?

I searched the thread and it would be nice if somene who successfully modded k10 => grid k2 can summarize how its done (in detail).

It would be a valueable addition to this thread to have the mod in written form even though there is this video on youtube.

thanks a lot!


edit: resistors as described in pic of post #1411 (first link of this post) works for grid k2 PCI ID ... next step will be the bios.

will update soon.

[RoGeorge]

Hello.

I checked this thread after many years but nothing changes (still replacing resistors on SPI lines  ).
A few years ago I challenged myself to enable GTX/RTX/Quadro cards for vGPU paravirtualization and I was successful.

Check this: https://gridforums.nvidia.com/default/topic/8934/

The presented solution is proving that there is no differences between the same chip in GTX-Quadro-Tesla lines and all is about "software".
The solution is useful for virtualization only! It enables the vGPU feature for compatible GTX/RTX/Quadro cards to TESLA/GRID counterparts like M10,M60,P4,T4. It does not modify host installed card (no HW mod, no vBIOS mod, no SW driver modification, probably EULA compliant). It also relaxes all NVIDIA "crippled/throttled" vGPU features to guests (like GPU MEM limits, number of emulated monitors (max 4), resolution of monitors, CUDA...). It does not remove NVIDIA license (to be compliant).

It's great to hear about successful mods.   

I've tried about two years ago to passthrough a GTX760 card with either Xen or Proxmox, and failed.  No idea if the fail was because of my lack of skills, lack of the proper SW patches, or the lack of proper hardware.  At most, Windows 7 was detecting the videocard, but nVidia drivers were refusing to install saying something about no nVidia GPU was present.

Later, I've read that only Quadro cards and alike can be video passtgrough, no matter what hypervisor is used (and GeForce cards will not work because of software limitations imposed by nVidia).

- Is it possible to passthrough a GTX760 without hardware modding?
- If not, does the method of changing the resistors applies to GTX760, too?
- Is the following hardware configuration enough for video passthrough, please?

• GPU:  GTX760 2GB (https://www.asus.com/Graphics-Cards/GTX760DC2OC2GD5/)
• CPU:  i7-4790K (https://ark.intel.com/content/www/us/en/ark/products/80807/intel-core-i7-4790k-processor-8m-cache-up-to-4-40-ghz.html)
• MB:  Z97 (https://www.asrock.com/mb/Intel/Fatal1ty%20Z97%20Professional/)

[colokarl]

I know KVM is able to passthrough any nvidia card to the guest.
It is capable of hiding the virtualization to the guest.

i did this succesfully many times with gtx 10xx cards.

[colokarl]

lol, i just saw mcerveny already mentioned kvm's capabilty to passthrough.

@mcerveny: i am intrigued by your "magic script" - out of interest, do you think your software mod could be transfered to esxi (for homelab)?

Plus i see you did use fedora with xen kernel, so it should be appliable to any distribution with xen kernel?

How long did it take to successfuly use your sw mod including research, writing the script ?

It is really f* up how nvidia rips of customers by selling basically the same hardware.

Appreciate some info on this - Thank you

[gotofbi]

lol, i just saw mcerveny already mentioned kvm's capabilty to passthrough.

@mcerveny: i am intrigued by your "magic script" - out of interest, do you think your software mod could be transfered to esxi (for homelab)?

Plus i see you did use fedora with xen kernel, so it should be appliable to any distribution with xen kernel?

How long did it take to successfuly use your sw mod including research, writing the script ?

It is really f* up how nvidia rips of customers by selling basically the same hardware.

Appreciate some info on this - Thank you

vGPU and pass through is different
I was able to use vGPU function with ESXi 6.5 without any issue after modding from K10 to K2
Only problem that Im still facing is compatability issue with Dell R720
It works fine on other workstation but not in R720

[colokarl]

Yes, of course vgpu!=passthrough even though both need vt-d capable mobo

My new homelab setup will be exactly like this... modded k10s => grid k2 running vgpu in esxi 6.5.

Can you confim windows 10 guests will install and run the k2 drivers, cause xenserver 6.2 seems not be able to do this as per videos from reply #1400 on page 57.

Right now i am waiting for the 3d printed active coolers for the modded k10s as i wont use 19" servers.

Also havent flashed k2 bios yet, but this will be easy.

Anyway, thank you for the reply and the valuable info that our beloved modded cards will run vgpu on esxi 6.5.

[colokarl]

Wow, so it took a while to get to the goal 

As i wrote in my previous post i will go for esxi as i did the resistor mod to my cards.

I tip my hat in regards to your software mod, well done !

Thank you, i do like it when someone discovers ways around vendor specific limitations.

 

[coolo]

mcerveny, hi
Your magic script is very interesting script for me, I have 1080ti, but want to try how work grid in rdp sessions...
I never used virtualization, but I think, that I can't write same this magic script, because I have not any grid card as etalon, for discover how it works and after put 1080ti and see any diffs for future script... Is it true?
Could you make this mod without grid card?
Thanks

[coolo]

The magic script is the last part of VDI puzzle.
(btw. RDP does not use accelerated GPUs until you unlock it in registry - https://lmgtfy.com/?q=rdp+use+accelerated+GPU+registry)

Yes, I tried it with my GTX 1080ti, my rdp server with 1 rdp session was PC with win10 pro, and slave was laptop with win 10 pro... I turned on in registtry acceleration but it doesn't work...
And I thought, that grid card can help me.... And easiest way to try it was your decision

[colokarl]

The magic script is the last part of VDI puzzle.
(btw. RDP does not use accelerated GPUs until you unlock it in registry - https://lmgtfy.com/?q=rdp+use+accelerated+GPU+registry)

Yes, I tried it with my GTX 1080ti, my rdp server with 1 rdp session was PC with win10 pro, and slave was laptop with win 10 pro... I turned on in registtry acceleration but it doesn't work...
And I thought, that grid card can help me.... And easiest way to try it was your decision

Hey coolo,

can you summarize what exactly you want to achieve and what your setup (harware operating system etc) is like ?

Maybe we can help you out better then.

cheers

[coolo]

Yes, sorry that I choose decision without description not my problem...

At the end I want get:
RDP host for work through thin client
Working 3D acceleration in RDP for work witch CAD application and rendering.

Now I have 2x1080ti. Work with CAD in obviously PC with cards.
But I want RDP

[colokarl]

Yes, sorry that I choose decision without description not my problem...

At the end I want get:
RDP host for work through thin client
Working 3D acceleration in RDP for work witch CAD application and rendering.

Now I have 2x1080ti. Work with CAD in obviously PC with cards.
But I want RDP

Alright,

Windows Server has its own hw-acceleration for vgpu called "RemoteFX", i havent used that yet, and i am not sure if the consumer cards like the gtx 1080 are supported:

You have to research if a gtx 1080 will work.

This should point you in the right direction:

https://github.com/MicrosoftDocs/windowsserverdocs/blob/master/WindowsServerDocs/virtualization/hyper-v/deploy/deploy-graphics-devices-using-remotefx-vgpu.md

Code: [Select]

Host system requirements:

    Windows Server 2016
    A DirectX 11.0-compatible GPU with a WDDM 1.2-compatible driver
    A CPU with Second Level Address Translation (SLAT) support

If this setup works, you will rdp to the server and use the hosts gpu for your cad applications installed on the host.

Also using hyper-v if you want to virtualize your cad boxes is possible (gpu passthrough)

Let us know how it goes...

[gotofbi]

@mcerveny

I am very interested your magic script
This is very exciting!! I didnt read few posts carefully and totally missed this golden information!

The presented solution is proving that there is no differences between the same chip in GTX-Quadro-Tesla lines and all is about "software".
The solution is useful for virtualization only! It enables the vGPU feature for compatible GTX/RTX/Quadro cards to TESLA/GRID counterparts like M10,M60,P4,T4. It does not modify host installed card (no HW mod, no vBIOS mod, no SW driver modification, probably EULA compliant). It also relaxes all NVIDIA "crippled/throttled" vGPU features to guests (like GPU MEM limits, number of emulated monitors (max 4), resolution of monitors, CUDA...). It does not remove NVIDIA license (to be compliant).

As you mentioned It does not remove NVIDIA license.
That means I still need to purchase GRID vPC license? ($50 per CCU subscription per year)

I really wanted to build home VDI infra with KVM support.
I was hoping K10 -> K2 mod will give me VDI in KVM but griddy nVIDIA do not support K1/K2 to work on KVM (not in grid but only with vgpu software)
Do you think its possible to bring K2 to support with vGPU5~10 software? or its better to start with vgpu software supported whatever GPU?

I was thinking about getting AMD S7150 for KMV support but if theres a hole in vgpu software, I really want to approach that way.

My ultimate goal is to run 16 VDI infra with Looking Glass(https://looking-glass.hostfission.com/)

[colokarl]

Maybe we shold crowdfound mcerveny to make his code open source and continue to develop on github as a team based project

Pretty sure this will p*ss off nvidia a lot and could trigger legal actions (?) but i'd also love to see the actual code, especially for homelab projects.

[gotofbi]

Maybe we shold crowdfound mcerveny to make his code open source and continue to develop on github as a team based project

Pretty sure this will p*ss off nvidia a lot and could trigger legal actions (?) but i'd also love to see the actual code, especially for homelab projects.

I spend some time of reading mcerveny's post in gridforum.
It looks like he is using vGPU with proper license.
His magic script allows user to use consumer graded card to vGPU with legit license.

Unless someone pirate/cracks the license file, his magic script will not be useful that much (unless get eval license every 90 days with new account )

[colokarl]

... true indeed.

i remember the sh*tstorm when nvidia released its new licensing model.

[gotofbi]

... true indeed.

i remember the sh*tstorm when nvidia released its new licensing model.

Indeed, it seems to be easy crackable model if I didnt misunderstand the concept.
It looks like download license file from nvidia and load into local license server (java)
And local license server gets communicated with all otherr VDI VM and controls license bs

Am I understandig it correctly?
It looks like local license server can be run without synced to nvidia license server as long as it has license.bin file

[colokarl]

... then it should be worth having a look at this.

My reverse engineering skills may be a lil rusty, but i'd be in for the lulz.

[gotofbi]

09:15:33,055 INFO  Starting FlexNet License Server 2018.09 (build 241024)
09:15:33,057 INFO  Copyright (C) 2013-2018 Flexera Software LLC.
09:15:33,057 INFO  All Rights Reserved.

They are utilizing 3rd party software to manage licenses.
 

[colokarl]

https://webcache.googleusercontent.com/search?q=cache:QiFwYJRxW0UJ:https://forum.exetools.com/showthread.php%3Ft%3D18864

I'll do some more research later, link contains some useful info.

@gotofbi : are you sure, there is no online check against nvidias backend ?

...because, unfortunately i think there is

[gotofbi]

https://webcache.googleusercontent.com/search?q=cache:QiFwYJRxW0UJ:https://forum.exetools.com/showthread.php%3Ft%3D18864

I'll do some more research later, link contains some useful info.

@gotofbi : are you sure, there is no online check against nvidias backend ?

...because, unfortunately i think there is

I only checked with nvidia's documentation.
I didnt see any mention that Virtual GPU Software License Server needs to be communicated with nvidia server
Moreover, the diagram doesnt say its communicating with nvidia server



Why do you think theres online check?

[gotofbi]

Yes, true. But you must upload license file (license.bin) within 96 hour from download time from licensing portal. The file is locked to first ethernet MAC address. But guest drivers now communicate with NVIDIA (one driver parameter from over 700  - "RMNvTelemetryCollection") and you must login to NVIDIA to use "Geforce experience"...

So theoretically, if I have license file with knowing generated date and mac address, I should able to install in any VM and continue to use it right? 
I checked guest driver configuration part and it looks like its only communicating with local license server.
I dont really care about "Geforce experience" as long as I can run DX application
However the license server is little overwhelming to reverse engineer it and Ill probably give up and move to MxGPU

[colokarl]

... thanks for clarifying mcerveny and yes, its offtopic.

To get back on topic:

I soon will flash the k2 bios to the modded k10s.

https://drive.google.com/drive/folders/1jp824XrO6fn1bPTCbrFjUEVBtxTptB1Z

I guess the 2014 Version is preferable for this or is there any reason i should go for the 2013 Version ?

@mcerveny: one last lic question:

Since its pretty easy to change the MAC Adress of eth0, have you tried using the same lic File on another Subnet, would that be worth a try ?

[friskydingo]

I'm trying to mod a Quadro K1000M to a Geforce GT 630M. I have it reporting as a 630M under device manager and GPU-Z but its still showing as a K1000M because the device id hasn't changed.Are there resistors on the K1000M that I can modify or can I change the id in the bios?

[gotofbi]

off topic,
With hint and help of mcerveny, I made it work on KVM.
I dont have magic script and its all accomplished with binary patch so I cant release.