[MOVED] Hacking NVidia Cards into their Professional Counterparts

[zizux]

I've saved the bios off a GTX 480, then changed the device ID with nibitor to a 06D9 from 06C0 but it still shows up with the old device ID after flashing.

nvflash -5 does require me to type YES due to mismatch but I'm doing something wrong.

EDIT FIXED. Flashed using the --straps and rebooted and now its working fine.

[gordan]

I've saved the bios off a GTX 480, then changed the device ID with nibitor to a 06D9 from 06C0 but it still shows up with the old device ID after flashing.

nvflash -5 does require me to type YES due to mismatch but I'm doing something wrong.

NiBiTor doesn't change the device ID strap, only the tag (the tag doesn't do anything, other than make nvflash refuse to flash the BIOS without the override). Next time please read and understand the writeups linked in this thread about how to carry out the modification before reducing the signal-to-noise ration on this thread.

[zizux]

Ah I see where I messed up and now have a working one.  Now on to setting up a lab to test it with.

Thanks!

[beckend]

So I had some problems with Tesla K10, I had the wrong memory size(29xxMB) and insufficient info in GPU-Z, and nonexistent sensors.
Maybe it suppose to be like that but drivers was kinda messed up too, got "no screen attached to the gpu" msg when going into Nvidia Control Panel.
So I removed the the 2nd resistor nibble also to make it a GRID K2, and now everything works as it should.

[gordan]

So I had some problems with Tesla K10, I had the wrong memory size(29xxMB) and insufficient info in GPU-Z, and nonexistent sensors.

Sounds like you didn't reinstall the driver properly after you modified the card.

[kdot]

Please,can anybody make a video on replacing resistors

[TheBay]

It would be nice if you would offer a service of modding the cards Some of us have unsteady hands. Only good for nvflash strapping ..

the best avice I can give is not to skimp on soldering iron. Get a good one with pointed tips so you can apply the tiniest of solder amounts.

Or go one better and get a hot air reflow iron. A decent one of those costs more than a Titan, though.

One other possible alternative is to get conductive glue.

But none of those preclude the requirement for very, very steady hands; a good night's sleep and no caffeine for 24 hours before attempting makes a noticeable difference.

I could easily solder anyone’s card if they were nearby me.

Those ultra pointed tips are a pain imo. I found that they loose heat when you touch the component and you have to have the heat turned up more. When its just sitting there it starts to oxidize unless you turn it back down. They just aren't a big enough heat sink and the temp control is all out of whack.

I use a hakko 936 with the 900M-T-B tip that came with it. I think they are around the $80 mark now. When I was working I preferred the chisel type D tips but they are easy enough with either. I used to do 0603s at work.

The key really is flux if you don't get it soldered right away before the flux in the solder evaporates. Get some no clean flux in a syringe tube and some fine tweezers.

My arms or hands shake a fair bit. I just hold them hard against the desk while I'm soldering.

I find the Weller WSP80 is just the ticket with the tiny tip on it for SMD work.

[ryanistheryan]

Update:
As some of you that have been following this thread are aware, I've been having a rather bizzare problem with only DL-DVI working on my modifed GK104 based cards (680 and 690). It just occurred to me that the two cards I have happen to have on thing in common - they are both Gainward cards. Has anyone else successfully managed to get DL-DVI to work in a VM with a Gainward GTX680 Phantom 4GB or Gainward GTX690? Gainward 690s have a few strapping resistors in different places to the EVGA and other 690s, so it is plausible they made some tweaks that cause the problem. Does anyone have either of those cards working with DL-DVI outputs successfully?

On a separate note, I just learned that GTX780Ti has device ID 0x100A. K6000 is 0x103A. We only need to change the 3rd nibble (via oguz286's awesome mod). His mod is to use a 33K resistor to jack up the 3rd nibble from 0 to 2. On my Titan using an 18K resistor instead boosts the ID to 3. So to make a 780Ti to a K6000, simply apply an 18K resistor between VCC and SCLK and voila, job done. For extra points, solder a couple of wires between those pins, solder an 18K resistor on one of them, and a switch to connect them. Break the switch out somewhere accessible (extra extra points for making the switch easily and neatly accessible from the back of the card without obstructing the airflow too badly). Now you can switch between a 780Ti and a K6000 at a flip of a single switch.

Needless to say, the neatness of this means that I'll be acquiring a 780Ti as soon as I've ebayed my Titan.

Edit: Maybe I'll keep my Titan for a little bit longer, as if the PCB is the same (which it almost certainly is), it should be very obvious which resistors are configured differently for the 4th nibble. Watch this space.

Hey there gordon, I have been lurking this thread on and off since it was discovered.

I modded a Zotac GTX 680 into a K10, with hopes that I could get it to do GPU accelerated computation in this program called CST. CST does not have support for any consumer video cards for acceleration.

After the mod, it shows up at as K10 in windows, but trying to use the video out of the card results in a limited resolution of 1280x720, and using programs such as CST or even attempting to benchmark it with other software results in it not being recognized. Do you believe this a hardware issue or a manufacturer issue?

[gordan]

I modded a Zotac GTX 680 into a K10, with hopes that I could get it to do GPU accelerated computation in this program called CST. CST does not have support for any consumer video cards for acceleration.

After the mod, it shows up at as K10 in windows, but trying to use the video out of the card results in a limited resolution of 1280x720, and using programs such as CST or even attempting to benchmark it with other software results in it not being recognized. Do you believe this a hardware issue or a manufacturer issue?

I have found that both my Gainward 680 and 690 only provide video output at SL-DVI modes from inside a VM (but they work fine with DL-DVI on bare metal). You'll find that on a typical monitor, SL-DVI only goes up to 1280x800. If you are having this issue on bare metal, however, that is unusual - I don't think anyone has seen that problem before. For example, my MSI 680 works fine, so I assumed this was a Gainward specific issue, since I seem to recall other people modifying Zotac cards without issues before.

Once you modified the card, does it reliably show up as a K10 on device manager and GPU-Z? Have you installed the Quadro driver (as opposed to the GeForce driver), and what version? I have been successfully using drover 320.92 Quadro driver. Does CST require a dedicated non-primary card for computation perhaps?

[mrkrad]

you know that 1280 bug is documented by xen as a known issue even with grid

[ryanistheryan]

Hey Gordon!

Thanks for the extremely quick reply, I know it isn't your job to provide support around this but I highly appreciate talking to another experienced user as I am still a student.

I have no issues with it popping up as a K10 with GPU-Z or Device Manager. I've used Quadro/Tesla drivers 307.45 (recommend version by CST) as well as the latest drivers available.  Have you had issues with different drivers? I am trying to reread all the threads from where I've left off here and there. I decided to test gpu acceleration with another program called cgminer (yeah I'm just dabbling with that stuff), and it correctly functions with my stock 560 ti, but thereafter says it says GPU 1 failure.

I have additions GPUs driving my screen, but as they were all nVidia based I didn't want to have driver conflict issues possibly, and that is when I realized the GTX680-modded-to-K10 would not go past 1280.

On page 17 verbigbad boy stated:
"I successfully modified
Zotac PCI-E NV ZT-60206-10L GT640 Synergy 2G 128bit DDR3 900/1600 DVI*2+mHDMI RTL
To NVIDIA GRID K1. It is working fine. passthough works too. BUT Device ID modification posible only after bios modification. Bios modification is needed only for specific vendors.
"

Do you think this would be an issue with this zotac? I'm going to search the thread for other zotac 680s done successfully. I'm not very proficient editing BIOSes and uploading them, I'd have to find a tutorial if that is the case. 

Side note, when looking through CST's guidelines for GPU acceleration, they show that the nVidia control panel in windows has a Tesla specific branch after stereoscopic 3D video settings. I don't have this available within my control panel.

Thanks for your help,

-Ryan

[mitiay]

So, I was able to mod the titan into the k20xm.

I need mostly DP precision in EM simultaion and want to do the same mod.
Am I correct: all that needs to be done is changing the resistor 25k->40k and flashing the modded 128kb BIOS from k20xm?

[gordan]

you know that 1280 bug is documented by xen as a known issue even with grid

Link to the ticket/errata?
I have pinned down the issue to SL-DVI vs. DL-SVI, and it only manifests on my Gainward cards GK104 based cards (e.g. my MSI 680 works fine). For example, if I use a SL-DVI monitor like IBM T221, I can run 3840x2400@16Hz over a single SL-DVI output, and that works fine from a VM.

I have no issues with it popping up as a K10 with GPU-Z or Device Manager. I've used Quadro/Tesla drivers 307.45 (recommend version by CST) as well as the latest drivers available.  Have you had issues with different drivers? I am trying to reread all the threads from where I've left off here and there. I decided to test gpu acceleration with another program called cgminer (yeah I'm just dabbling with that stuff), and it correctly functions with my stock 560 ti, but thereafter says it says GPU 1 failure.

I have additions GPUs driving my screen, but as they were all nVidia based I didn't want to have driver conflict issues possibly, and that is when I realized the GTX680-modded-to-K10 would not go past 1280.

On page 17 verbigbad boy stated:
"I successfully modified
Zotac PCI-E NV ZT-60206-10L GT640 Synergy 2G 128bit DDR3 900/1600 DVI*2+mHDMI RTL
To NVIDIA GRID K1. It is working fine. passthough works too. BUT Device ID modification posible only after bios modification. Bios modification is needed only for specific vendors.
"

Do you think this would be an issue with this zotac? I'm going to search the thread for other zotac 680s done successfully. I'm not very proficient editing BIOSes and uploading them, I'd have to find a tutorial if that is the case. 

Side note, when looking through CST's guidelines for GPU acceleration, they show that the nVidia control panel in windows has a Tesla specific branch after stereoscopic 3D video settings. I don't have this available within my control panel.

If BIOS modification were required, the card wouldn't have showed up as a K10 in device manager / GPU-Z, but it is easy enough to check. Download the bios using nvflash -b, and check the block verybigbadboy mentioned. You will also want to check the more recent Kepler UEFI header BIOS strap discussion more recently on the thread since there are only 5 device ID controlling bits in the strap which could be relevant to the device ID changing, and you don't want to disrupt the rest.

Modified cards don't get the extra features showing up in the control panel. It is not entirely clear at the moment why, but some of it is due to the BIOS. For example, flashing a K5000 BIOS onto a modified 4GB GTX680 will make the ECC control show up, but other features are still missing.

Also note that AFAIK you can't use the TCC mode on your primary display adapter.

So, I was able to mod the titan into the k20xm.

I need mostly DP precision in EM simultaion and want to do the same mod.
Am I correct: all that needs to be done is changing the resistor 25k->40k and flashing the modded 128kb BIOS from k20xm?

It may not work - you need to modify the 4th nibble value to match the Tesla card, and that hasn't been located yet. The drivers check the hard strap rather than the soft strap to enable features. Titan allegedly has full DP performance, but my Titan shows near identical figures for DPFP in CUDA-Z as my 780Ti, so either CUDA-Z isn't measuring it reliably or it is a myth that Titan has uncrippled DP performance.

The only full ID modification available at the moment is 780Ti to K6000, since those don't differ in the 4th nibble, so changing only the 3rd nibble is sufficient.

Also note that video outputs and configuration are set by the BIOS, so flashing a Tesla BIOS onto a GeForce card will disable all video outputs on the card.

[mitiay]

It may not work - you need to modify the 4th nibble value to match the Tesla card, and that hasn't been located yet. The drivers check the hard strap rather than the soft strap to enable features. Titan allegedly has full DP performance, but my Titan shows near identical figures for DPFP in CUDA-Z as my 780Ti, so either CUDA-Z isn't measuring it reliably or it is a myth that Titan has uncrippled DP performance.

The only full ID modification available at the moment is 780Ti to K6000, since those don't differ in the 4th nibble, so changing only the 3rd nibble is sufficient.

Also note that video outputs and configuration are set by the BIOS, so flashing a Tesla BIOS onto a GeForce card will disable all video outputs on the card.

Read the thread once again and found that the resistor mod which seemed to work is 33k between VCC and SCLK.

According to previous unlucky experiments, resistors controlling the 4th nibble should be located on the card's backside against the flash, right?
By the way, if softmod doesn't work then why original Titan BIOS contains FF FF FF 7F 00 14 00 80 at 0x1C, which, if I'm not mistaken, changes hardware coded 4th nibble from something (0, 1, 4 or 5) to 5?

Regarding the Titan's DP uncrippled performance, AFAIK it is activated by the driver which changes the card's BIOS so it turns on all DP FPUs and at the same time drops frequency to 732 MHz. Do you know if anyone made a comparison of "crippled" and "uncrippled" BIOSes?

[gordan]

Read the thread once again and found that the resistor mod which seemed to work is 33k between VCC and SCLK.

To modify the 3rd nibble from 0 to 2. To modify it to 3 you need an 18K resistor.

According to previous unlucky experiments, resistors controlling the 4th nibble should be located on the card's backside against the flash, right?

That's what I thought but comparing a Titan to a 780Ti revealed no differences; specifically, none of the resistors are in alternate locations as one might expect given the ID differences.

By the way, if softmod doesn't work then why original Titan BIOS contains FF FF FF 7F 00 14 00 80 at 0x1C, which, if I'm not mistaken, changes hardware coded 4th nibble from something (0, 1, 4 or 5) to 5?

Not on all Titans. On mine there are no overrides. The soft strap will affect which driver gets installed (it affects the device ID), but it won't affect the feature set. Kepler GPUs seem to have a register that gets set by the hard strap, regardless of what the soft-strap says. If you change the soft-strap without changing the device ID, use GPU Caps Viewer and see what it says - you'll find that it reports a device ID as set by the hard strap, even though the device ID is overriden by the soft strap. The driver looks at whatever stores the hard strap value to enable/disable various features.

Regarding the Titan's DP uncrippled performance, AFAIK it is activated by the driver which changes the card's BIOS so it turns on all DP FPUs and at the same time drops frequency to 732 MHz. Do you know if anyone made a comparison of "crippled" and "uncrippled" BIOSes?

No idea. I'm running a fully delimited engineerion BIOS on my Titan, but I have no immediate use for DP performance so haven't investigated in detail.

[beckend]

So I had some problems with Tesla K10, I had the wrong memory size(29xxMB) and insufficient info in GPU-Z, and nonexistent sensors.

Sounds like you didn't reinstall the driver properly after you modified the card.

Actually I tested the card on 2 of my computers, I would be satisfied with the TESLA but clean the drivers multiple times and even reinstalled windows on my main rig would not help. However it worked instantly when modded to the GRID K2, and I have no idea why.

[gordan]

Strange. Had you verified that it is coming up with the correct device ID in GPU-Z and GPU Caps Viewer?

[kwikit]

Gordan,

I wanted to say you've been doing a heck of a job passing on the information. Kudos for that.

I ran into this site over the weekend, read all 57 pages, and am willing to try a few things on an EVGA GTX 780 SC w/ ACX.  Problem is, I'm not at all electronically savvy, so I need a little guidance.

I'm building 8 graphics workstations, and chose the 780 because some people in the group wanted to use Iray in 3dsmax, which uses the CUDA cores to render when present.

I was disappointed to learn, the 3D viewport performance in 3ds max 2014 is horrendous with these cards. To give you a baseline, the card in my current machine which is a Dell T1500 i7-870, with an AMD Radeon 4800, performs as well and marginally better than the 780 using the same exact scene / settings for comparison. I tested with an FX 580 as well, and those are 5 times slower than both the 4800 / 780. I suspect there is amazing performance available in the 780 if it can be unleashed, and the proper Quadro settings enabled.

Before I came to this site, I did edit the inf files, to no avail, and was researching soft modding the card, but quickly learned that likely won't lead anywhere. In my research, I found TechPowerUp, which had some hi-res photos of the bare video card sans any coolers.  I don't know if this helps, or if it's 100% accurate as some of the traces are quite faint especially around the EEPROM, but I lightened up the file in Photoshop and created a PDF with the trace lines highlighted. I think it's interesting there are 4 main trace lines going to the processor.  I haven't yet taken the time to open up my own card since I want to be sure I have the parts / tools to test / assemble the solution. When I do, I will try to take some higher res shots. http://www.techpowerup.com/reviews/EVGA/GTX_780_SC_ACX_Cooler/3.html

I'm inclined to add on resistors to convert the 780 to a Quadro 6000. I know, 18K resistor between SCLK / VCC...  I need to at least find the 4th nibble, but I'd also like to modify the 3rd nibble at the resistor and add on an appropriately sized resistor in parallel if that's viable.

What do i need to get started in testing? Multimeter? Variety of resistors to test, or a rheostat to try various resistors? But once I have a working solution with a rheostat, can i use the multimeter to determine the right resistor value to install permanently?  Given the size, I'm assuming i can order some 1/4w resistors of various ohms and that will be a good kit to work with, or that rheostat i mentioned.  If you could link me to a suitable multimeter / resistors and give me a few hints on how to proceed, I'll do the testing and report back.

What I need to know is, to test, I assume the card is installed in the slot and the PC powered up, so I can take measurements?  I'm mostly concerned about frying the card before I get the answer.
I plan to create a bootable USB Dos Disk, so I will use NVFLASH to check what value I'm getting as I attach / detach resistors or adjust the rheostat.

Just so you know, I think the 4 traces define the ID, and I think the Red trace is the 4th nibble, but I won't know till I try. I'm hoping I can get hold of everything by Christmas, so I have something to do when i get bored. :-)

Thank you for taking the time to read even if you can't help answer. I'll sort it out one way or the other.

[kwikit]

I ended up buying these as they looked to be the appropriate tools.

http://www.amazon.com/gp/product/B000EVYGZA/ref=oh_details_o00_s00_i00?ie=UTF8&psc=1

http://www.amazon.com/gp/product/B003UC4FSS/ref=oh_details_o00_s00_i01?ie=UTF8&psc=1

[gordan]

@kwikit

As has been mentioned on the thread before, viewport performance in 3DSMax appears to not be affected by the GPU much. It is single threaded and mostly CPU bound. ATI drivers are a little thinner which is why they have better performance in this application. I tested a genuine Quadro 2000 against a GTS450, and the same GTS450 modified into a Q2000, and there was on observable difference in viewport performance.

I am not sure if those resistors you bought are going to be suitable (I'm not sure they will fit under the big metal sandwich plate that covers the PCB) - but I guess it depends on what exactly you are trying to do. Are you planning to use them for a remote switch setup on a 780Ti like I described, to be able to toggle it between a 780Ti and a K6000?

[kwikit]

Gordan,

I read your assessment but I figured it was worth a test on 1 card, if i can do it before the others need to be exchanged for real quadro's. My understanding was that the quadro's do perform better in viewports than the mainstream cards but I am not sure I want to make that investment to find out. We actually picked up a quadro k4000 to test, just haven't opened it.

I know it's directx 11 that's driving the viewports, but I was under the impression that nvidia had optimized some code which is enabled on the quadro's to work through directx better. My aim is to try to get that feature enabled in the gtx 780. One way towards that aim is to make the driver believe the gtx is the quadro. Maybe it won't make any difference in which case, I may just end up pairing the gtx with a cheap lightweight gt or gts or other card that can run the viewports adequately and have the gtx as a coprocessor of sorts for the CUDA bit. Not what I want to do, so I may rethink the Quadro K4000 vs GTX 780 decision. I just don't want to open a $850 Quadro K4000 to find out it equally sucks in the viewports... then I'd want to scream bloody murder.

If the ID works out as I hope, I can always buy the right resistors... this pack contains a broad range of values so I grabbed it. By the way, I think my diagram of the traces is missing some interconnections, the image is playing light tricks, so I won't know for sure till i open my own card, but looking at Guz's site, his image seems to show a few more interconnects. And it makes sense one could connect the SCLK with VCC based on the traces... and I wonder if the other nibble can be tracked down by looking for a similar connection to a power line.
http://www.guztech.nl/wordpress/index.php/2013/11/researching-nvidia-gpus-geforce-gtx780-and-gtx-titan-to-tesla-k20-and-tesla-k20x/

So anyone relying on my marked up image, be aware there are probably some missing connections and maybe even some erroneous connections shown in that diagram.

I am not sure I needed the multimeter, but it doesn't hurt having one in the toolkit. :-)

 

[gordan]

I read your assessment but I figured it was worth a test on 1 card, if i can do it before the others need to be exchanged for real quadro's. My understanding was that the quadro's do perform better in viewports than the mainstream cards but I am not sure I want to make that investment to find out. We actually picked up a quadro k4000 to test, just haven't opened it.

I haven't seen any evidence of real Quadros performing obviously better on viewports, but if you find otherwise, please, do post some benchmark numbers that demonstrate the difference with a repeatable test.

If the ID works out as I hope, I can always buy the right resistors... this pack contains a broad range of values so I grabbed it. By the way, I think my diagram of the traces is missing some interconnections, the image is playing light tricks, so I won't know for sure till i open my own card, but looking at Guz's site, his image seems to show a few more interconnects. And it makes sense one could connect the SCLK with VCC based on the traces... and I wonder if the other nibble can be tracked down by looking for a similar connection to a power line.
http://www.guztech.nl/wordpress/index.php/2013/11/researching-nvidia-gpus-geforce-gtx780-and-gtx-titan-to-tesla-k20-and-tesla-k20x/

For the 3rd nibble, a 1206 surface mount resistor is by far the best option - it is exactly the right size, and looks tidy.

IIRC johnjoe experimented with the resistors on the EEPROM side of the PCB and didn't find any that affect the 4th nibble. I did a visual check of the back of the PCB between a 780Ti and a Titan, and couldn't spot any resistors in that area that were in alternate locations.

[trr444]

guys, i did conversion of the 780ti, it recognizes as K6000 both in device manager, nvidia control panel and cuda-z but still there is no "workstation" context menu in the nvidia control panel.. i downloaded k6000 drivers fron nvidia but it did not help. i need mosaic to set up. also cuda-z performance is same in 780ti and k6000 modes, should be there any difference? whats wrong? please advice.

[gordan]

There is no expected performance difference from modifying. You get TCC mode (of it's a secondary card) and VGA passthrough, and that's about it. I don't know about Mosaic, I only use XP and that doesn't support Mosaic anyway. Have you tried downloading the Mosaic app from Nvidia? I seem to remember you need that to make it work - if it is going to work. I seem to recall that gnif originally mentioned something about better multi-display support under Linux, but I never found I needed it as Linux/Xorg comes pretty well equipped to deal with multiple displays anyway (my primary Linux Xen dom0 card is an unmodified GeForce card running a T221 in dual-input mode).

[trr444]

thanks for reply. I did not find any mosaic tool, it is not available for download... NVIDIA manual says it is included in the NVIDIA Control Panel and once quadro card is in the system Mosaic settings become available in the Workstation context menu.I need Quardo to setup several  monitors as a single big screen, this is that Mosaic offers. Also, i tested performance with the SPECviewperf, and Catia-03 test show 23 mean value, that is a bit low for K6000.current results for even K5000 is 75. You sure card is working in Quadro mode and not just showing it is Quadro?