[MOVED] Hacking NVidia Cards into their Professional Counterparts

[vgeek]

For you who like to try these cards in vSphere and Horizon View 5.2, there is a new whitepaper from VMware that explains how to configure both vSGA (Shared GPU) and vDGA (Dedicated GPU):

http://www.vmware.com/files/pdf/techpaper/vmware-horizon-view-graphics-acceleration-deployment.pdf

[maxlapshin]

Is it possible to download nvidia bios under linux?

[gordan]

Bah. I just modified my GTX580 into a Quadro 7000 only to find out that my main/only reason for modifying it (gaming in a VM) is not applicable on this card - Quadro 7000 is not supported for VGA passthrough! 

[verybigbadboy]

Bah. I just modified my GTX580 into a Quadro 7000 only to find out that my main/only reason for modifying it (gaming in a VM) is not applicable on this card - Quadro 7000 is not supported for VGA passthrough! 

Hello
You may try to change device id to quadro 6000. I think it is strange to mod 1-GPU GeForce card to 2-GPU Quadro.
Also can you provide picture with resistors positions please ?

Thank you

[gordan]

You may try to change device id to quadro 6000.

That was what I looked into first, but the device ID between 580 and 7000 can be adjusted by only twiddling the bottom 5 bits o fthe ID. To go all the way to the 6000 requires modifying the bottom 13 bits, which is more difficult.

I think it is strange to mod 1-GPU GeForce card to 2-GPU Quadro.

You mean like modifying a GTX680 to a Grid K2?
Anyway, Quadro 7000 is a single-GPU card. I suspect you are confusing Quadro 7000 with QuadroPlex 7000. The latter is an external PCIe enclosure with two Q7000 cards in it. A Q7000 card is essentially identical spec to a GTX580, apart from having 2-4x the RAM on-board (6GB on Q7000 vs. 1.5-3GB on a 580).

Also can you provide picture with resistors positions please ?

I'll see what I can do, but I lack a decent digital camera to do this with, and my cheap ass Android phone's camera barely has enough dots to make people's faces vaguely recognizable (as in recognizable as human), let alone surface mount components.

[verybigbadboy]

You may try to change device id to quadro 6000.

That was what I looked into first, but the device ID between 580 and 7000 can be adjusted by only twiddling the bottom 5 bits o fthe ID. To go all the way to the 6000 requires modifying the bottom 13 bits, which is more difficult.

You found one of resistors, It's very good May be I try to find second one?
I think it is possible to find photo via google, can you provide your card name please?

I looked bios from 580. It is different than 600 series bios but I don't see any softs straps in bios. Looks like "and" and "or" masks allow any values. If I right It means you need only hardware mod

thank you

[gordan]

After my failure to get a GTX580 to do what I want, I got my hands on a cheap GTS450 (thanks for your investigation on those cards earlier on the thread). Successfully modified it to a Q2000.

Good news: Works fine for gaming in a Xen VM.
Bad news: There is absolutely no performance boost in SPECviewperf from modifying them to Quadros. I have a genuine Quadro 2000 and that utterly annihilates the modified GeForce card in SPECviewperf.

The thing is - I cannot figure out why. I went as far as flashing a Q2000 BIOS to a GTS450, hacked about with the BIOS, compared clocks, GPU-Z and CUDA-Z info, and absolutely everything I can see shows that the GTS450 should be about 50% faster because it has the same number of shaders and memory and is clocked about 50% higher.

What is the special ingredient that makes the drivers decide to use a non-crippled OpenGL renderer? They clearly identify the card as a quadro, otherwise it wouldn't work in the VM. Most perplexing...

[gamezr2ez]

What is the special ingredient that makes the drivers decide to use a non-crippled OpenGL renderer? They clearly identify the card as a quadro, otherwise it wouldn't work in the VM. Most perplexing...

I suppose I should ask. You are running the Quadro drivers, yes?

Also, try these performance tests under native windows, not through Xen. You have to do some wierd stuff to get full options under Xen.

Bah. I just modified my GTX580 into a Quadro 7000 only to find out that my main/only reason for modifying it (gaming in a VM) is not applicable on this card - Quadro 7000 is not supported for VGA passthrough! 

Nether is the K5000. It still works. I would suggest using the Geforce drivers and checking that. When modding to the Quadro on my card, it had different results when using the Geforce vs Quadro drivers.

[gordan]

What is the special ingredient that makes the drivers decide to use a non-crippled OpenGL renderer? They clearly identify the card as a quadro, otherwise it wouldn't work in the VM. Most perplexing...

I suppose I should ask. You are running the Quadro drivers, yes?

Of course.

Also, try these performance tests under native windows, not through Xen. You have to do some wierd stuff to get full options under Xen.

I am running them on bare metal Windows.

Bah. I just modified my GTX580 into a Quadro 7000 only to find out that my main/only reason for modifying it (gaming in a VM) is not applicable on this card - Quadro 7000 is not supported for VGA passthrough! 

Nether is the K5000. It still works. I would suggest using the Geforce drivers and checking that. When modding to the Quadro on my card, it had different results when using the Geforce vs Quadro drivers.

I uninstalled the GeForce drivers - they are newer and clobber the Quadro drivers. WEIRD things happen when you have a GeForce and a Quadro in the same system. But putting in a Q2000 on it's own makes SPECviewperf go fast. Pulling it and replacing it with a modified GTS450 makes it go slow. Same driver, same version, according to device manager.

[gordan]

Quick question guys - regarding the GTX690 mod - can the ID of both GPUs be modified? I'm pondering unifying two gaming rigs into one. Get a GTX690, pass one GPU to each VM. One big box, 2 monitors, 2 keyboards, and 1x GTX690->Grid K2 seems like a neater solution than 2x GTX680s (that and I'm a little short PCIe slots for other things need in there).

[free_electron]

What is the special ingredient that makes the drivers decide to use a non-crippled OpenGL renderer? They clearly identify the card as a quadro, otherwise it wouldn't work in the VM. Most perplexing...

fusebits in the core. disabled. you think NVidia is so stupid that it would simply be a few resistors on the outside ?

[amigo]

@gordan

I am in the same boat, I have a GTX 570 and found the necessary resistors for it to go to Q7000 and Tesla C2075, C2090 but there's no performance changes.

What have you tried doing to the GTX BIOS to adjust it to the new mod?

Flashing the card to the actual Quadro/Tesla BIOS does not work because of timings and other things, but these parameters can be used as transplants.

I would be interested if you have worked further on getting the upper 13 bits figured out for the Quadro 4/5/6000, if that's even an option?

[amigo]

fusebits in the core. disabled. you think NVidia is so stupid that it would simply be a few resistors on the outside ?

That's an interesting hypothesis, but do you have anything to support it?

[free_electron]

That's an interesting hypothesis, but do you have anything to support it?

Here's how it works :

The resistors are used to downgrade boards depending on what is installed. You can only go down , not up. The max capability of the GPU is set in the GPU. The resistors determine the downmix (memory, speed, cores, voltage ,core voltage, ram voltage) .

You can set resistors to a combination above what the core is specced at but it will not enable. it is a logical OR of what the resistors say and what the fusebits say. Fusebits have priority. Sometimes there are no fusebits as the chip simply does not have the extra features. One board layout can be used for multiple different GPU's in a family.

[gordan]

@free_electron - not always the case.

In the case of GTS450 vs. Quadro 2000, however, I believe I have just figured it out - it is actually a slightly different chip between the two; I missed a tiny detail in the spec readings.

GTS450 == GF106
Quadro 2000 == GF106GL

I'm guessing the "GL" is important. 

Never mind - I got the VM VGA passthrough functionality for free - that was the main (and only) thing I was after anyway.

[amigo]

Here's how it works :

The resistors are used to downgrade boards depending on what is installed. You can only go down , not up. The max capability of the GPU is set in the GPU. The resistors determine the downmix (memory, speed, cores, voltage ,core voltage, ram voltage) .

You can set resistors to a combination above what the core is specced at but it will not enable. it is a logical OR of what the resistors say and what the fusebits say. Fusebits have priority. Sometimes there are no fusebits as the chip simply does not have the extra features. One board layout can be used for multiple different GPU's in a family.

Would you happen to know which features would be disabled in the core though that relate to Quadro capability?

Or is that that the fusebits only manage the features that exist in all chips from the same family (ie. GF100 or GF110: GTX480/GTX570 vs Quadro 4-5-6000/Quadro 7000). Therefore if the fusebit is set, the driver disables the feature but the feature still exists in the chip itself.

[SeanB]

To do a zener zap on bits during an on wafer test that does a basic function test that then determines whether the chip is fully good, usable for the top line, or partially good and usable with disabling some cores ( you will have a little logic to enable selective disabling of cores and then reoranising the remainder to appear as contiguous) so it will be a lower spec chip, or is a DOA or fails in critical areas ( same thing really, if the cache logic or glue logic is faulty then it will never work properly) so is recycled, and these pads are never bonded out to pads on the BGA is cheap, and might only add 1 second more to the initial test, at the lowest cost part of the process, before you add an expensive package to what will be a dead unit. Then you can slice and sort based on the paint dot on the chip, and grade them into the bins as needed for assembly. Then you take the complete units and place in a test jig for speed grading, and then you can use resistors or links to enable features. That way you will get different chip suffixes, as they are now different items. If you are lucky they ran out of the low spec chips as they were doing the run for your card and then finished off with the full spec units, as they do cost the same. That though is the luck of the draw, you could have 2 cards made on the same line at the same day but they used different chips on it.

[amigo]

Though question still remains, what does make Quadro QUADRO?

There are no special fancy features between the two lines of chips, perhaps slight core difference, if any, such as Shaders / TMUs / ROPs. But that explains the above process of selection so really the features are all there. So what decides whether something is a Quadro and what does Quadro do more that GeForce doesn't.

What I mean is that if you run a SPECViewPerf you will get two different results between the same line of chips (ie. GF100) just because one is GeForce and the other is Quadro. What cripples the performance, OGL driver, BIOS, hardstraps, fuses,...?

[jimseal]

Trying to do a proof of concept before investing in the real thing.  I have an ASUS GTX660TI-DC2T-2GD5 card and would like to mod it to a Quadro K5000 or some other GPU virtualization compatible card.  (Hyper-V 2012 ?).  The ASUS card seems different from the other images in this thread.  Here are some pictures of both sides, if someone would be so kind as to help me out:

Thank you,

[verybigbadboy]

Trying to do a proof of concept before investing in the real thing.  I have an ASUS GTX660TI-DC2T-2GD5 card and would like to mod it to a Quadro K5000 or some other GPU virtualization compatible card.  (Hyper-V 2012 ?).  The ASUS card seems different from the other images in this thread.  Here are some pictures of both sides, if someone would be so kind as to help me out:

Thank you,

Hello, looks like they rotated GPU
Can you take a picture of the area shown on a pic?

I googled some photos, and I think I found locations of resistors.

[free_electron]

@free_electron - not always the case.

like i said : same pcb layout can hold multiple chips. some are actually different , some are fusebits.

A key element of Quadro is the drivers. if you got a regular card and you experience a software glitch nobody gives a rat's ass.
If you have a quadro and there is a glitch with a particular software , it will get fixed and it will get fixed very quickly.
quadro's are workstation card. meaning they are used to visualised data from engineering software. solidworks, catia , adobe .. all the big stuff out there.

there are other elements on quadro . OpenGL is run in hardware on a quadro s opposed to software emulation on a non quadro. Quadros can have much more memory than other boards  (which is important for massive polygon scenes . And no SLI does not share memory ... ) Try loading a complete Catia design for a wing of an airplane down to the last nut and bolt. change the viewport. The Quadro will respond immediately. the GTX will curl up in a corner ,choke and may give you a screen refresh a few minutes later. unworkable if you are designing.

As software is being rewritten to take advantage of CUDA the quadro will loose its advantage. but tright now , 3d cad software is , in a lot of cases, still driven on OpenGL or DirectX and there the quadro runs circles around the other boards.

in short : for gaming , video processing : no advantage , stick with gtx
for anythin involving massive vertex count : quadro will chomp the gtx to bits. so if you are thinking to make the next Toystory or Monsters Inc , or run things like 3Dmax, Catia and other high end cad software ... quadro is the way to go. games have no advantage on quadro as the amount of polygons a game throws at the GPU is simply too small to notice an effect.

[jimseal]

More info on the 660Ti ASUS

Hope that helps...

Thanks

[verybigbadboy]

More info on the 660Ti ASUS

Hope that helps...

Thanks

Thank you for perfect photos

You may use resistors value table from 680

[sw]

Hello, I have a EVGA GeForce GTX 660Ti  (03G-P4-3663-KR) that I would possibly like to try this on, I have read through this forum but still have a few questions. It seems to me this card can be shown as a tesla k10, quadro k5000 or a grid k2 according to a post on page 16. I game a lot, but also use processor intensive programs like Adobe CS6 and a few others that would seem to benefit from these higher end cards. Which would be the best card for me to change the resistor values to, and will I loose PhysX support or anything similar by reporting this cards to the computer as a higher model?

[gordan]

... will I loose PhysX support or anything similar by reporting this cards to the computer as a higher model?

I have found this to be somewhat hit-and-miss. GTS450 -> Q2000 appears to lose PhysX on XP-64, but not on Win7-64. GTX480->Q6000 seems to retain PhysX capability on XP-64. The only way to really know for sure is to try it and see, over and above going over the spec of all quadro cards on the Nvidia website and checking which ones they claim support PhysX.