[MOVED] Hacking NVidia Cards into their Professional Counterparts

[verybigbadboy]

Can someone help me? I am trying to make a k5000 from my MSI gtx 660 ti, but I don't know where I can find the resistors and what I need to change them with.

Thanks!

Hello - I was taking a look at the photos that you posted however they aren't all that clear. Try to get some better lighting and try and get the tiny details, lettering etc.. all in focus - this will really help in getting identified what needs to be identified

hq scan. same board.
http://www.ixbt.com/video3/images/msi-7/msi-gtx660ti-scan-front.jpg
http://www.ixbt.com/video3/images/msi-7/msi-gtx660ti-scan-back.jpg

[vacaloca]

Interesting, so Titan remains the only one with uncrippled DP FP.
It makes sense, I suppose - they had to sacrifice something to keep the GPU with the extra few shaders enabled from cooking itself at gaming grade clocks and voltages required.

Having said that - what about modding the 780Ti into a Titan? If Tom's Hardware is correct and it is due to the driver lowering the DP FP clock speed, the modding it into a Titan would work around this and give you the best of both worlds - extra shaders and full DPFP performance.

Well, almost the best of both worlds, heh. You'd still be missing TCC support. One part that still seems to be missing is how to enable full DPFP performance on conversions to cards that support it natively (e.g. Quadro K6000) Titan is just the only card (so far) that is able to enable/disable full DPFP on the fly, so presumably the 780Ti -> Titan conversion would permit full DP support. Would be interesting if someone tried it!

[johnjoe]

So, I was able to mod the titan into the k20xm. I attached a pic of cuda-z. The titan values are without activated double precision values. The k20xm is running with activated double precision. Nevertheless, the performance of the k20xm is not exactly what I expected. I increased already the clock, but it wasn't the solution. By the way, the mod of oguz286 is correct (resistor). There must something else to increase the performance. But the advantage in speed is already noticeable in EM-simulations. So far, maybe somebody has an idea how to improve the performance!

--Update--
I noticed one interesting thing: nvidia-smi reports, clock is running with 732 MHz (TITAN 835MHz or something). If I modify the bios with kepler bios tweaker it doesn't change at all. Cuda-z reports for example 900Mhz but the performance doesn't change. Maybe nvidia-smi is correct and I have to change something else.

[vacaloca]

--Update--
I noticed one interesting thing: nvidia-smi reports, clock is running with 732 MHz (TITAN 835MHz or something). If I modify the bios with kepler bios tweaker it doesn't change at all. Cuda-z reports for example 900Mhz but the performance doesn't change. Maybe nvidia-smi is correct and I have to change something else.

You're doing it right if nvidia-smi reports maximum clocks of whatever maximum clocks you modified it with. Of course this might vary since I don't believe K20xm has any notion of Boost clocks -- but if you are still running the original Titan BIOS, then it might actually make a difference:

I had to change the Max GPC, L2C, XBA, SYS Boost States table on my (real) Quadro K6000 BIOS to the clock I set on the Common tab in Kepler BIOS Tweaker for nvidia-smi to report the correct maximum clocks and see the effect on my CUDA code.

[gordan]

Is the performance difference between a vanilla Titan and the one modified to a Tesla coming from TCC mode or from something else?

Note that on Kepler class GPUs soft-modding doesn't _actually_ work properly. It sort of almost pretends to work (as in nothing complains), but it doesn't enable any of the extra features. See my post earlier about this - it looks like the driver gets the hard-strap values (check what GPU caps tool reports as device ID vs. the PCI device ID) and acts accordingly. I will know for sure tomorrow when my GTX 690 is fully modified. The key point being that you likely need to hard-mod the 4th nibble as well. If I am reading the diagram correctly, that would mean hooking up a resistor of currently unknown value between VCC and SO on the EEPROM. I have not attempted this yet, so do so at your own risk.

[johnjoe]

Actually, the 4th nibble is hard strapped with a res of 33k between vcc and sck, as posted by oguz286. I think the dp performance is ok, only single precision is poor

[gordan]

Actually, the 4th nibble is hard strapped with a res of 33k between vcc and sck, as posted by oguz286. I think the dp performance is ok, only single precision is poor

That's the 3rd nibble, not the 4th.

Are you using the standard Titan BIOS, or did you flash the Tesla K20c BIOS onto it?

[johnjoe]

Ah, ok you mean i have also to adjust the 4th nibble. It is not a titan or k20c bios. It is a extracted k20xm bios from ibm.

[gordan]

Ah, ok you mean i have also to adjust the 4th nibble.

Indeed. Soft modding it _might_ work for this, but I have a sneaky suspicion that I'm about to find out that hard-modding is fully required on Keplers, possibly due to Nvidia getting smart to soft-modding. So not only did they not extend the soft-mod strap bits to cover more than the lowest 5, but there is a good chance that the GPU itself remembers the ID as hard strapped, before the EEPROM chip re-straps it with the soft-strap. It's a hunch, so take with a pinch of salt for now. As I said, I'll know in a couple of days when my 690 is back and fully hard-strapped.

It is not a titan or k20c bios. It is a extracted k20xm bios from ibm.

Oh, I see. Any particular reason for that? It might be interesting to try and see if you get different performance with different BIOSes (assuming clocks and power budgets are all adjusted to the same Titan levels using Kepler BIOS Tweaker).

Note: there has recently (a few days ago, on the IRC channel) been some work going on on decoding the power/voltage tables within the nouveau project, and it is unclear whether Kepler BIOS Tweaker might be partially incorrect in it's understanding of these. Those tables are not really used by the BIOS itself, they are there for the driver to parse out and deal with accordingly under various load conditions, hence nouveau project interest in it.

[vacaloca]

Actually, the 4th nibble is hard strapped with a res of 33k between vcc and sck, as posted by oguz286. I think the dp performance is ok, only single precision is poor

Can you be more explicit? are you saying that DP performance isn't crippled and is 1/3rd the SP performance in your conversion?

[gordan]

Actually, the 4th nibble is hard strapped with a res of 33k between vcc and sck, as posted by oguz286. I think the dp performance is ok, only single precision is poor

Can you be more explicit? are you saying that DP performance isn't crippled and is 1/3rd the SP performance in your conversion?

My understanding is that Titan's DP performance is uncrippled anyway (it was one of it's key selling points, being placed between the GeForce and Quadro/Tesla cards in this regard), and johnjoe is modding a Titan (unless I'm misunderstanding what's going on, which may well be the case). Unlocking the full DP performance on a 780/780Ti, OTOH, would be big news.

[IanFist]

[...] I have a 4GB GTX680 running a K5000 BIOS, and there is no obvious performance benefit in any test over a standard GTX680 - the only advantage is in persuading the drivers to allow VGA passthrough operation - and even that doesn't seem to work on models more recent than the GTX680.

Hey gordan,
Could you tell me if changed anything in the bios before flashing it? Or could you just flash it after you changed the hard straps?
Where did you obtain this bios?

oguz286, you are a legend! Any chance of a before/after BIOS hex diff? I'd rather like to try to flash a GTX480 with a Q6000 BIOS with RAM size adjusted appropriately and see what effect it has.

Also did you get any further with this?

I think I have a different motivation than most users here. It seems Nvidia have crippled 3Dsmax Viewport performance in their recent cards severely. Since 3Dsmax does not use any of the advanced OpenGl features of the Quadros but Direct 3D (9.0) I don't see any reason why geforce cards should perform as bad as they do. Especially the new kepler cards. A gtx780 is performing alot worse than a gtx 480. We have also have a few gtx 680s and the newest one performs as poorly as the gtx780 although it has the same specs as the older ones.
A Quadro K2000m performs alot better then any of the geforce cards.

I first tested with a gtx480 but changing the sof straps and flashing it with the hardware id to a Quadro 6000 but did not see any performance gains in 3dsmax.

[Brechreiz]

Hi everyone,

thank you for having this delightful topic on the site and all the contributions.
I just registered as i wanted to ask if anyone having a a quadro mod could confirm that the quad buffer function for openGL is available and functional.
If someone has the kindness to spend some minutes for testing, please open the nvidia control panel, goto "manage 3D settings" and find "Stereo enable" on the list. Set it to "enable" as well as the stereo output mode to "generic shutter" (or sth like this) and reboot your computer.
Afterwards the quad buffer should be available.
I dont want to post direct links to downloads - but i have a 50kb exe that then could simply check if the QB is really enabled and available. I've attached a screenshot of the app running and stating that QB is enabled.

Also, I know I'm a first time poster without any reputation.
I've uploaded the QB Test to virustotal and here are the results:
https://www.virustotal.com/de/file/2beb32433c74cbac757c5f63e06562884db0a33714d3bc39a44ea8cdb373f207/analysis/1385032172/

Thank you for your time!

Kindest regards
Brechreiz

[gordan]

[...] I have a 4GB GTX680 running a K5000 BIOS, and there is no obvious performance benefit in any test over a standard GTX680 - the only advantage is in persuading the drivers to allow VGA passthrough operation - and even that doesn't seem to work on models more recent than the GTX680.

Could you tell me if changed anything in the bios before flashing it? Or could you just flash it after you changed the hard straps?
Where did you obtain this bios?

Just the straps.
I can't remember where I got it, most likely from the TPU BIOS database.

oguz286, you are a legend! Any chance of a before/after BIOS hex diff? I'd rather like to try to flash a GTX480 with a Q6000 BIOS with RAM size adjusted appropriately and see what effect it has.

Also did you get any further with this?

No, it's a lot of tedious work to analyze the binary init script, and I am not convinced the result would be worthwhile.

I think I have a different motivation than most users here. It seems Nvidia have crippled 3Dsmax Viewport performance in their recent cards severely. Since 3Dsmax does not use any of the advanced OpenGl features of the Quadros but Direct 3D (9.0) I don't see any reason why geforce cards should perform as bad as they do. Especially the new kepler cards. A gtx780 is performing alot worse than a gtx 480. We have also have a few gtx 680s and the newest one performs as poorly as the gtx780 although it has the same specs as the older ones.
A Quadro K2000m performs alot better then any of the geforce cards.

I first tested with a gtx480 but changing the sof straps and flashing it with the hardware id to a Quadro 6000 but did not see any performance gains in 3dsmax.

As far as I can tell from my limited testing, the 3DSMax viewport is software rendered. I tried a genuine Quadro 2000, GTS450, GTS450 modified into a Q2000, GTX470, GTX470->Q5000, GTX480, GTX480->Q6000, GTX580, GTX580->Q7000, GTX680, and GTX680->K5000 and I got about 7fps for the viewport out of all of them, no difference between them at all. Which implies software rendering.

[gordan]

General update:

1) Verified that soft-modding DOES NOT WORK on Kepler class GPUs (GTX6xx and later). It _looks_ like it works - the PCI device ID changes. But if you look at the GPU caps viewer, it shows the hard strap device ID. That is the ID the driver keys off, and if you don't set the full hard strap correctly, the driver will load and pretend to work, but will treat the card as a normal GeForce card and will not allow VGA passthrough to work. I have just confirmed this as the key reason why my GTX690/K2 was not working.
1.1) When you part-soft-mod, the driver will initialize the PCIe speed at 1.1 only, regardless of what the card and the bus are capable of. This is the reason why oguz286' mod was causing PCIe speed to always read v1.1 rather than higher.

In conclusion to this part - don't bother soft-modding the straps on the Keplers, it is a waste of time and will just have you chasing your tail for ages trying to figure out why everything seems to check out but doesn't work.

oguz286, if you're reading this, to get your PCIe speed above 1.1, hard-mod the 4th nibble on your 780 and you'll find it magically speeds back up. You might find performance improves in other things, too.

2) Not all GTX690s are the same. I have re-created the problem KamranB had with the modified GTX690 not working. The problem is in the mod for the 3rd nibble (if you want to mod to K2, just leave the 4th nibble resistor off). If you set that resistor to 20K, it seems to cause the SCLK output like going to the GPU to get attenuated too far and all you get is 0s, which will mean the EEPROM cannot be read. That means the card doesn't get it's BIOS initiaized and although it'll show up on the bus, it won't work. Same thing happens if you do a similar mod to oguz's VCC-SCLK resistor and use a resistor that has too low a value. You can leave the 3rd nibble resistor off but the 3rd nibble value will flap between 0xA and 0xB. I have slightly stabilized mine to the point where it is usable if you don't mind sometimes having to double-reboot by putting a 33K resistor across VCC-SCLK, but this is without any of the 3rd nibble resistors fitted. There is probably a better way to do this (maybe 40K on the original 3rd nibble resistor to replace the default 25K, or a lower value instead of 20K on the alternative location for it). I haven't experimented with this more because 1206 component soldering is not far off my manual precision limit, and 0402 is just most certainly beyond it.

GTX690 3rd nibble resistor findings
2.1) 20K on alternative location attenuates too high, eeprom becomes unreadable, card doesn't initialize, and cannot be re-flashed (don't try to re-flash it in this state, erase will work but writing will not, returning a 256-byte page programming error).
2.2) Without either 3rd nibble resistor, 15K across VCC-SCLK attenuates too high and results in the same situation as 2.1)
2.3) Without either 3rd nibble resistor, 27K across VCC-SCLK yields device ID flapping between 0x119F () and 0x11AF and 0x11BF. Haven't quite figured out what is going on here.
2.4) Without either 3rd nibble resistor, 33K across VCC-SCLK yields reasonable stable device ID of 0x11BF. Sometimes it flaps down to 0x11AF, but much less frequently than it would without this. Maybe a 40K resistor across those two pins would fix the problem, but I don't have any handy to try.

Anyway, the most important point here is that the documented mod for the 690 isn't universally applicable, so bear this in mind if you are modding a 690.

3) On my system the 690 is having the same dual-link mode problem as my 680. It works fine with SL-DVI modes, but causes the monitor to go to lose signal and go to sleep with a DL-DVI mode is used. Deeply bizzare and I have no explanation available for this one at all, but that does appear to be what happens. I would like to stress that this does not happen on my Quadrofied 480 (Q6000), DL-DVI modes work just fine on that. But since it is happening to me on two different cards modded in different ways, it is probably safe for others to disregard as it seems specific to my system for some unfathomable reason (but if you find the same happens to you, please do let me know, we might be able to figure out what is going on by comparing notes.

4) I have figured out the byte that controls the hierarchy ID (PCIe switch port on multi-GPU cards like the 690) in the BIOS. This would allow you to flash, say, a K5000 BIOS onto the 690, but this wouldn't work because you would need to figure out which bit in the BIOS init sets the memory size and halve it for the K5000 BIOS to work on the 690 (or a 2GB 680).

5) Because of this bizarre DL-DVI issue my 690 isn't going to be useful to me, so if somebody wants it for further experimentation, modding or just using, make me an offer in line with what they sell for on ebay. It's the Gainward model.

And now back to modding the Titan into a K6000. Need to figure out if I can pull the 4th nibble up via the SO line.

[suntzu]

I think I have a different motivation than most users here. It seems Nvidia have crippled 3Dsmax Viewport performance in their recent cards severely. Since 3Dsmax does not use any of the advanced OpenGl features of the Quadros but Direct 3D (9.0) I don't see any reason why geforce cards should perform as bad as they do. Especially the new kepler cards. A gtx780 is performing alot worse than a gtx 480. We have also have a few gtx 680s and the newest one performs as poorly as the gtx780 although it has the same specs as the older ones.
A Quadro K2000m performs alot better then any of the geforce cards.

I first tested with a gtx480 but changing the sof straps and flashing it with the hardware id to a Quadro 6000 but did not see any performance gains in 3dsmax.

As far as I can tell from my limited testing, the 3DSMax viewport is software rendered. I tried a genuine Quadro 2000, GTS450, GTS450 modified into a Q2000, GTX470, GTX470->Q5000, GTX480, GTX480->Q6000, GTX580, GTX580->Q7000, GTX680, and GTX680->K5000 and I got about 7fps for the viewport out of all of them, no difference between them at all. Which implies software rendering.

You forgot about CPU limit. There is a small sweet point where 3ds Viewport is not CPU limited...
...Sometimes in 3ds "simpler" GPU is better.
1scenario: you have simple scene few low poly objects. CPU can't generate enough data to fill GPU. So you can mod as you want... CPU will be limited. What GPU will be quicker? Simply fermi with 4SM=>192 shaders load  50%, 770 8SM->1536 shaders load with 8%?

Logic choice would be to make complex scene with advance textures and high poly objects so we create scenario 2:
Some highpoly objects, and a lot low poly like flowers, trees etc.
Again it can be CPU limited scene, a lot Call Draws make CPU bound since GPU will have to rendering a lot of big/small things CPU have to generate a lot of commands to GPU.

Since SandyBridge -> Haswell we gain about 8-20% of performance, GPU we get more than 2 times more power. Always have been that sometimes we were CPU bound, sometimes GPU bount. Now in most cases we are CPU bound.


same here

[IanFist]

As far as I can tell from my limited testing, the 3DSMax viewport is software rendered. I tried a genuine Quadro 2000, GTS450, GTS450 modified into a Q2000, GTX470, GTX470->Q5000, GTX480, GTX480->Q6000, GTX580, GTX580->Q7000, GTX680, and GTX680->K5000 and I got about 7fps for the viewport out of all of them, no difference between them at all. Which implies software rendering.

hey gordan, thanks for answering my questions.
There are different viewport rendering modes available actually. Since 3dsmax 2012 the default is the Nitrous renderer which should be states it is directx9 based ( directx 11 for max 2014). On top of that fps changed for me when changing to different cards. surprisingly best results came from quadro k2000m ( but admittedly different system) and radeon 5870.

for the comparison a large scene with high polycount and complex materials was used.

But even considering sun tzus (which I didn't fully understand considering my findings) post I still cant find an explanation why the newer gtx680 in my testing only returned about half the fps from the older one.

[gordan]

As far as I can tell from my limited testing, the 3DSMax viewport is software rendered. I tried a genuine Quadro 2000, GTS450, GTS450 modified into a Q2000, GTX470, GTX470->Q5000, GTX480, GTX480->Q6000, GTX580, GTX580->Q7000, GTX680, and GTX680->K5000 and I got about 7fps for the viewport out of all of them, no difference between them at all. Which implies software rendering.

hey gordan, thanks for answering my questions.
There are different viewport rendering modes available actually. Since 3dsmax 2012 the default is the Nitrous renderer which should be states it is directx9 based ( directx 11 for max 2014). On top of that fps changed for me when changing to different cards. surprisingly best results came from quadro k2000m ( but admittedly different system) and radeon 5870.

for the comparison a large scene with high polycount and complex materials was used.

But even considering sun tzus (which I didn't fully understand considering my findings) post I still cant find an explanation why the newer gtx680 in my testing only returned about half the fps from the older one.

It depends on how the rendering is coded. I have seen a number of reports saying that a GTX480 and GTX580 (i.e. top of the line Fermi, there is very little difference between the 480 and 580) significantly outperforms GTX680 in some tasks. Crysis fps and shader counts aren't the only things that define performance. That's why the Quadro 6000 (GF100, spec between GTX470 and GTX480) still fetches almost as much on ebay as the Quadro K5000 (GK104, same spec as GTX680).

[suntzu]

As far as I can tell from my limited testing, the 3DSMax viewport is software rendered. I tried a genuine Quadro 2000, GTS450, GTS450 modified into a Q2000, GTX470, GTX470->Q5000, GTX480, GTX480->Q6000, GTX580, GTX580->Q7000, GTX680, and GTX680->K5000 and I got about 7fps for the viewport out of all of them, no difference between them at all. Which implies software rendering.

hey gordan, thanks for answering my questions.
There are different viewport rendering modes available actually. Since 3dsmax 2012 the default is the Nitrous renderer which should be states it is directx9 based ( directx 11 for max 2014). On top of that fps changed for me when changing to different cards. surprisingly best results came from quadro k2000m ( but admittedly different system) and radeon 5870.

for the comparison a large scene with high polycount and complex materials was used.

But even considering sun tzus (which I didn't fully understand considering my findings) post I still cant find an explanation why the newer gtx680 in my testing only returned about half the fps from the older one.

So simply... when viewport is CPU bounded better performance came from better driver, architecture GPU, not horse power.
If you are CPU limited let's assume that you need 50% power of 480...
What will be faster? 480 or Titan?
Ofc 480. Since 240 CC@1,4GHz can get job done much faster than 500 CC@750MHz. Fermi can be TWICE faster than titan if whole frame could be count at one cycle.
If GPU get job done twice faster, CPU can clean things and prepare next frame.

It's know true that AMD driver is much less CPU hunger driver than NV. That's a reason why Radeon 5870 could be faster than NV card's in DX
Another problem here is NV cripple OpenGL performance of GF more than AMD in my experience.... you used DX viewport:)

[gordan]

Servus,

I changed the marked resistor in the pic from 25k to 40k. Now the pci-id is 1025 instead of 1005 for a Titan. The aim is to get 1020 (K20X). I changed already the resistors near by. But it doesn't changed anything. Somebody any ideas?

Did you have a chance to do any more investigation on the 4th nibble hard-strap on the Titan?
If it works the way I think it does (and that's a big if), the 4th nibble is controlled by the resistor pair directly to the left of the one you changed to 40K to boost the 3rd nibble to 0x2 (0x1005 -> 0x1025). I think that in order to get to 0x1020 you would need to replace the existing resistor in that pair to the left to 5K. That is pulling to ground on the right side and seems to be connected to SO on the left side.

If I'm right, you should also be able to apply a similar trick to bridging VCC and SCLK with a resistor by connecting SO and GND. If the current ID is 5 that would make the resistor to be modded 30K. to reduce that overal resistance to 5K you would need a 6K resistor between SO and GND.

[mosmo]

It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

thanks

[ilya80]

Hi guys!

Just here to report we successfully removed resistors 1 & 3 from the reference design PNY GTX 680 and it now reports itself as GRID K2. Here's screenshot taken in bare-metal Windows 7 with Quadro/Grid/Tesla drivers installed:



Card model is:

http://www3.pny.com/GTX-680-2048MB-PCIe-font-colordc0431-BONUS-16ft-Smart-Active-High-Speed-HDMI-Cable-includedfont-P3123C506.aspx

Resistors removed are 1 & 3, picture below ( courtesy of the guys who did the mod in the first place )

[gordan]

If it works the way I think it does (and that's a big if), the 4th nibble is controlled by the resistor pair directly to the left of the one you changed to 40K to boost the 3rd nibble to 0x2 (0x1005 -> 0x1025). I think that in order to get to 0x1020 you would need to replace the existing resistor in that pair to the left to 5K. That is pulling to ground on the right side and seems to be connected to SO on the left side.

If I'm right, you should also be able to apply a similar trick to bridging VCC and SCLK with a resistor by connecting SO and GND. If the current ID is 5 that would make the resistor to be modded 30K. to reduce that overal resistance to 5K you would need a 6K resistor between SO and GND.

I just tried this and can confirm that this is not the case. It looks like the SO line is not what controls the 4th nibble. I also tried connecting HOLD# (directly connected to VCC, it seems) to SI with a resistor, and that didn't affect the 4th nibble either. Since those 3 pins (SCLK, SI and SO) are what is connected to the 3 resistor pairs near the EEPROM, it looks like SI and SO aren't what controls the 4th nibble. This also tallies up with johnjoe's findings, which means the other ID straps must be somewhere else. There are at least 5 pairs of plausible looking resistors on the back of the Titan PCB, right opposite the EEPROM. It is likely one of those (GTX680 has a similar arrangement, where the pull-down resistor controlling the 4th nibble is on the back but the pull-up is on the front).

3 of the pairs seem to be reading 0 resistance across the fitted resistors, so it probably isn't those. The remaining two pairs are highlighted in the attached photo, and one of each pair does seem to be connected to VCC and GND, as the strap resistors would be.

I don't suppose anyone here with a Titan or GTX780 happens to have some higher precision equipment and more dexterity than me to test that hypothesis?

[gordan]

It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

You'll find that even most non-reference GTX670/GTX680 cards only differ minimally from the reference design, and the strap resistors are in the same locations. I have a Gainward Phantom GTX680 card which is technically non-reference, and I successfully modified it. You could actually try a part-mod. If you want to use it for virtualization, I read somewhere that Tesla K10 is supported for PCI passthrough, which means you wouldn't even have to remove the resistor controlling the 3rd nibble - only remove the one controlling the 4th. That should give you ID 0x118F for Tesla K10 and you might find it works just fine. Best of all, the resistor that controls the 4th nibble is on the back of the card, which means you wouldn't even have to take off the heatsink. Please report back if/when you do it.

[mosmo]

It's hard not to get excited over modding a GTX to a Quadro or K series pro card. However, is there any particular reference or non-reference GTX 6 series that can be modded 100%? I am willing to take the risk if it's not too difficult.

Actually, I have the GTX670-DC2-4GD5 non reference, has anyone attempted it on this card and had any luck? Please let us know.

You'll find that even most non-reference GTX670/GTX680 cards only differ minimally from the reference design, and the strap resistors are in the same locations. I have a Gainward Phantom GTX680 card which is technically non-reference, and I successfully modified it. You could actually try a part-mod. If you want to use it for virtualization, I read somewhere that Tesla K10 is supported for PCI passthrough, which means you wouldn't even have to remove the resistor controlling the 3rd nibble - only remove the one controlling the 4th. That should give you ID 0x118F for Tesla K10 and you might find it works just fine. Best of all, the resistor that controls the 4th nibble is on the back of the card, which means you wouldn't even have to take off the heatsink. Please report back if/when you do it.

This is awesome advice, thanks. Can you help with the identification of the resistor(s)? I will take a pic of the card when I get home. Also, do you know if SLI will work if I pass through both cards to a VM?