[MOVED] Hacking NVidia Cards into their Professional Counterparts

[decaf]

Is it possible to do this with a kernel hack? By just changing I.D. So system and also nvidia's driver see Quadro I.D. and behave that way.

[gnif]

Is it possible to do this with a kernel hack? By just changing I.D. So system and also nvidia's driver see Quadro I.D. and behave that way.

I wish, I tried this first, the binary blob interrogates the card directly ignoring the kernel reported ID

[Neo_Moucha]

Is it possible to do this with a kernel hack? By just changing I.D. So system and also nvidia's driver see Quadro I.D. and behave that way.

this was possible in the past using "softquadro" with RivaTuner software, see forum.guru3d.com
but nvidia blocked it in newer drivers...

[gnif]

Is it possible to do this with a kernel hack? By just changing I.D. So system and also nvidia's driver see Quadro I.D. and behave that way.

this was possible in the past using "softquadro" with RivaTuner software, see forum.guru3d.com
but nvidia blocked it in newer drivers...

No, it was not modding the kernel, it was overriding the software straps by writing to the register on the GPU directly.

[moisyes]

Have you looked at the front of the PCB?

The same 8-pin soic that is under the resistors you changed is on the lower-right of the other GPU and surrounded by a few resistors and unpopulated spots. It looks pretty promising to me.

Agreed, but as stated, this could be completely unrelated to it, it could be controlling voltage and modding could potentially kill the GPU.

Do you see the straps here, anywhere?

http://www.ixbt.com/video3/images/gf110/gtx580-scan-back.jpg

I wonder what a GTX580 can become (if anything useful)

Nothing stands out, they may be on the front of the card.

Gnif, would it be possible to follow the tracks from the second processor pins the same way the are routed to the resistors to find them anyway?
I have a GTS 450 too and after searching the web a lot, couldn't find any hardmod information, I think for obivious reasons. Dou you have any information on how I could hack my 2GB GTS 450 into a Quadro one. I use linux and need to hardmod instead to softmod, what have abundant windows information. Thak you in advance.

[Neo_Moucha]

this should be theoretically possible - change it into Quadro 2000 but only for older models before 2011? (GF106 based)
I have one real Quadro 2000 in my possession, so I can scan it and we can look for differences...

[ihasmario]

Does this apply to lesser cards such as the 670?

Superb work.

[mungewell]

Have you looked at the front of the PCB?

The same 8-pin soic that is under the resistors you changed is on the lower-right of the other GPU and surrounded by a few resistors and unpopulated spots. It looks pretty promising to me.

Agreed, but as stated, this could be completely unrelated to it, it could be controlling voltage and modding could potentially kill the GPU.

I'd agree that the placement of the links is approximately correct. Given the coding (resistor values) this would be another hint that you've found the right ones.

Given that you've confirmed the resistors which control GPU1, you could X-ray the board to find out the ball that these go into and then trace from that ball on the other GPU - just looking at surface layers might give you enough.

If anyone has a dead card you could even 'heat gun' the GPUs off to buzz out tracking.

Anyway, congratulations on your find. Look forward to using the 670 as a Quadro for some future 3D work under Linux ;-).
Simon

[gnif]

Gnif, would it be possible to follow the tracks from the second processor pins the same way the are routed to the resistors to find them anyway?
I have a GTS 450 too and after searching the web a lot, couldn't find any hardmod information, I think for obivious reasons. Dou you have any information on how I could hack my 2GB GTS 450 into a Quadro one. I use linux and need to hardmod instead to softmod, what have abundant windows information. Thak you in advance.

Not easy, they are a 6+ layer PCB.

Does this apply to lesser cards such as the 670?

Superb work.

Read back through the thread to find the answer.

this should be theoretically possible - change it into Quadro 2000 but only for older models before 2011? (GF106 based)
I have one real Quadro 2000 in my possession, so I can scan it and we can look for differences...

That would be interesting, but it may yield nothing as the resistors may not be moved, but be different values.

I'd agree that the placement of the links is approximately correct. Given the coding (resistor values) this would be another hint that you've found the right ones.

Given that you've confirmed the resistors which control GPU1, you could X-ray the board to find out the ball that these go into and then trace from that ball on the other GPU - just looking at surface layers might give you enough.

If anyone has a dead card you could even 'heat gun' the GPUs off to buzz out tracking.

Anyway, congratulations on your find. Look forward to using the 670 as a Quadro for some future 3D work under Linux ;-).
Simon

Thanks, and yeah, If I had an X-ray machine I would have already done it . I am able to remove the chip with my SMD workstation, but I do not have a dead card to try it on, and it is not likely to get one as everyone that has a dead card will still have warranty as they are still a pretty recent card.

[ihasmario]

Read back through the thread to find the answer.

Are you interested in (borrowing) a 670 to test your suspicions you mentioned? I am local enough (WA though), and my 670 struggles to run my screens because they're incapable of NVIDIA mosaic, so it won't be missed too dearly if it goes wrong.

[gnif]

Read back through the thread to find the answer.

Are you interested in (borrowing) a 670 to test your suspicions you mentioned? I am local enough (WA though), and my 670 struggles to run my screens because they're incapable of NVIDIA mosaic, so it won't be missed too dearly if it goes wrong.

I would be willing to have a go at the card, but you must understand there is a risk to damaging the card in order to find the correct straps. And if it is possible, what would you want it to become? A 680, Quadro K5000 or Tesla K10?

[ihasmario]

Read back through the thread to find the answer.

Are you interested in (borrowing) a 670 to test your suspicions you mentioned? I am local enough (WA though), and my 670 struggles to run my screens because they're incapable of NVIDIA mosaic, so it won't be missed too dearly if it goes wrong.

I would be willing to have a go at the card, but you must understand there is a risk to damaging the card in order to find the correct straps. And if it is possible, what would you want it to become? A 680, Quadro K5000 or Tesla K10?

I don't have a preference between quadro or Tesla, because I assume both can perform mosaic. Essentially the problem is that I use T221s, which always identify as two screens, meaning that true multi-monitor support is impossible with NVIDIA consumer cards.

I'll send a PM.

[gnif]

Read back through the thread to find the answer.

Are you interested in (borrowing) a 670 to test your suspicions you mentioned? I am local enough (WA though), and my 670 struggles to run my screens because they're incapable of NVIDIA mosaic, so it won't be missed too dearly if it goes wrong.

I would be willing to have a go at the card, but you must understand there is a risk to damaging the card in order to find the correct straps. And if it is possible, what would you want it to become? A 680, Quadro K5000 or Tesla K10?

I don't have a preference between quadro or Tesla, because I assume both can perform mosaic. Essentially the problem is that I use T221s, which always identify as two screens, meaning that true multi-monitor support is impossible with NVIDIA consumer cards.

I'll send a PM.

Wow, that is an impressive LCD, I am still using 3x Dell 2100FP from 2005. As for Mosaic, I have not checked if a Tesla works with it, as the Tesla K10 never had video outputs on it, so not sure. IMO it would be better to go for the Quadro K5000 as it better meets the requirement and less likely a NVidia update would break video output on modded Tesla cards.

[amyk]

Given that you've confirmed the resistors which control GPU1, you could X-ray the board to find out the ball that these go into and then trace from that ball on the other GPU - just looking at surface layers might give you enough.

I was thinking of this too. Someone here has xrayed some boards, maybe you could ask her for advice?

[emoose]

Hi there, thanks for this

I just wrote up a pretty detailed post about how this thread made me look at my 3GB GTX 660 Ti's BIOS and find it was similar to a GTX 670s (and different to 2GB GTX 660 Ti's), but because I couldn't read the CAPTCHA I pressed request another, which refreshed the page and lost the post 

Basically my 660 Ti's BIOS is almost the same as a 670's: it uses the same board and SKU numbers (20040005) but has extra code inserted which uses the normal 660 Ti board number (20040001, the 670 has similar code at a different address, but uses the 20040005 board number instead), maybe to emulate/downgrade it to a 660?

After seeing this I pressed further and compared 670 2GB vs 670 4GB, and then mapped the values that were different onto my 660 Ti's BIOS (the addresses were a bit different but it wasn't hard to locate them), they matched with the 670 4GB

This started to make me think they might have just crippled 4GB 670s into 3GB 660 Ti's, until I opened up my card and found 6x2Gb chips (H5GQ2H24AFR, which is 1.5GB? Maybe I read the datasheet wrong, or the rest were on the back...) There were 2 unfilled spaces though, so I'm guessing it hasn't got the full 4GB

The datasheet of those chips mentions that they're 256-bit, but the 660 Ti is reported as only being 192-bit... Maybe flashing the 670 BIOS over would enable the full bandwidth? I'd be willing to try it but I'm worried that the BIOS might do a check against the hardware device ID... I'd guess not since you can change the HW and it still works, but that could be down to a combined Quadro/Tesla/Geforce BIOS... Any info about this would be great!

Also any info about bad flashes would be great too, the only things I can find about them are from BIOS modders, not crossflashing I'm scared that the wrong RAM config/HW device ID/other stupid check might throw off the card from even being detected in nvflash...

If anyone wants to look further:

660 Ti 3GB BIOS: http://www.techpowerup.com/vgabios/127140/EVGA.GTX660Ti.3072.120806.html
660 Ti 2GB BIOS: http://www.techpowerup.com/vgabios/127242/EVGA.GTX660Ti.2048.120910.html

670 2GB BIOS: http://www.techpowerup.com/vgabios/125688/EVGA.GTX670.2048.120807.html
670 4GB BIOS: http://www.techpowerup.com/vgabios/126722/EVGA.GTX670.4096.120712.html

[gnif]

Hi there, thanks for this

I just wrote up a pretty detailed post about how this thread made me look at my 3GB GTX 660 Ti's BIOS and find it was similar to a GTX 670s (and different to 2GB GTX 660 Ti's), but because I couldn't read the CAPTCHA I pressed request another, which refreshed the page and lost the post 

Basically my 660 Ti's BIOS is almost the same as a 670's: it uses the same board and SKU numbers (20040005) but has extra code inserted which uses the normal 660 Ti board number (20040001, the 670 has similar code at a different address, but uses the 20040005 board number instead), maybe to emulate/downgrade it to a 660?

After seeing this I pressed further and compared 670 2GB vs 670 4GB, and then mapped the values that were different onto my 660 Ti's BIOS (the addresses were a bit different but it wasn't hard to locate them), they matched with the 670 4GB

This started to make me think they might have just crippled 4GB 670s into 3GB 660 Ti's, until I opened up my card and found 6x2Gb chips (H5GQ2H24AFR, which is 1.5GB? Maybe I read the datasheet wrong, or the rest were on the back...) There were 2 unfilled spaces though, so I'm guessing it hasn't got the full 4GB

The datasheet of those chips mentions that they're 256-bit, but the 660 Ti is reported as only being 192-bit... Maybe flashing the 670 BIOS over would enable the full bandwidth? I'd be willing to try it but I'm worried that the BIOS might do a check against the hardware device ID... I'd guess not since you can change the HW and it still works, but that could be down to a combined Quadro/Tesla/Geforce BIOS... Any info about this would be great!

Also any info about bad flashes would be great too, the only things I can find about them are from BIOS modders, not crossflashing I'm scared that the wrong RAM config/HW device ID/other stupid check might throw off the card from even being detected in nvflash...

If anyone wants to look further:

660 Ti 3GB BIOS: http://www.techpowerup.com/vgabios/127140/EVGA.GTX660Ti.3072.120806.html
660 Ti 2GB BIOS: http://www.techpowerup.com/vgabios/127242/EVGA.GTX660Ti.2048.120910.html

670 2GB BIOS: http://www.techpowerup.com/vgabios/125688/EVGA.GTX670.2048.120807.html
670 4GB BIOS: http://www.techpowerup.com/vgabios/126722/EVGA.GTX670.4096.120712.html

No worries . I do not think that the ram configuration is stored in the BIOS at all as it is needed to be known before the GPU even reads the BIOS from the EEPROM. In earlier versions it was based on the hard straps, I do not see any reason why they would have changed this.

As for the RAM size, from what I can see, that module is 2Gb which is 0.25GB per chip * 6 = 1.5GB total. This is very odd unless I am also reading it wrong if you say that the card should have 3GB of RAM. Can you have a real close look at the card to be doubly sure that the part number you provided is correct? Also, you did count the chips on both sides of the PCB?

As for seeing cards with less ram accessible then is physically installed, I highly doubt this would ever occur, the cost saving to the mfg is too large to just disable/hide/waste the additional RAM.

[emoose]

No worries . I do not think that the ram configuration is stored in the BIOS at all as it is needed to be known before the GPU even reads the BIOS from the EEPROM. In earlier versions it was based on the hard straps, I do not see any reason why they would have changed this.

Well that's good news then , I was scared of flashing the 4GB BIOS in case it messed up because of the missing 1GB, and the 2GB one had me scared because of some different values in spots that matched between the 3GB 660 and 4GB 670, if that's true then the 4GB one is sounding even better

As for the RAM size, from what I can see, that module is 2Gb which is 0.25GB per chip * 6 = 1.5GB total. This is very odd unless I am also reading it wrong if you say that the card should have 3GB of RAM. Can you have a real close look at the card to be doubly sure that the part number you provided is correct? Also, you did count the chips on both sides of the PCB?

I took some pictures of the chips while I had it open:

Haven't checked the other side yet because of some metal grille covering it, I can see some sort of thermal paste coming through the holes of it though. I'll take it out later and have a check for definite.

As for seeing cards with less ram accessible then is physically installed, I highly doubt this would ever occur, the cost saving to the mfg is too large to just disable/hide/waste the additional RAM.

Ah yeah that's true, probably should have thought of that before I took it apart

Think I might wait and buy another 3GB 660 Ti when I get money, don't really want to go without if it messed up
If you get hold of a 670 can you note down the resistors? Might come in handy if the BIOS does some sort of check.
Edit: nevermind, just saw that it would be 5K and 10K, bit slow today
I'll pull the card out and take a peek under the grille now

Edit2: Yep, the other 6 were there, along with another 2 unfilled spaces Oh well, heres some pictures:
http://imgur.com/a/TolyJ#8
Now check out
http://images.anandtech.com/doci/5818/GeForce_GTX_670_B-1.jpg
Look similar? 
This is a normal 2GB 660 Ti (best pic I could find): http://rwlabs.com/images/articles/evga/660ti_superclock/08.JPG
Seems like this board might have a good chance

Edit3: Looking around led me to this spot for the resistors:

Which seems to have a different config to the 670: http://www.ixbt.com/video3/images/ref/gtx670-scan-back.jpg
I admit I'm no good when it comes to hardware though, so I'm probably wrong about this... Don't really have any idea how I'd find out where they are, this was just me checking for similarity

Edit4: Remembered you saying about when it's pulled in a different direction it's a different set of values...

90% sure this is the spot now, hope you can tell us more

[nhp12345]

Hi there,
Nice job gnif 
Could you show me your modified card`s benchmark results with 3DS Max, 3DMark... or some game benchmarks?!  I really want to see how it performs.
Thanks a lot and good luck in your further tweaking! 

[gnif]

Hi there,
Nice job gnif 
Could you show me your modified card`s benchmark results with 3DS Max, 3DMark... or some game benchmarks?!  I really want to see how it performs.
Thanks a lot and good luck in your further tweaking! 

I would love to when I get some more time. The only place you may see an improvement in performance is CUDA stuff as only one core has been modified. Once the other GPU is modified the results will be much more interesting.

[amyk]

Edit4: Remembered you saying about when it's pulled in a different direction it's a different set of values...

90% sure this is the spot now, hope you can tell us more

Either by coincidence or just reuse/modify of an existing PCB, the ID resistors seem to be always next to one of the heatsink mounting holes.

[gnif]

Edit4: Remembered you saying about when it's pulled in a different direction it's a different set of values...

90% sure this is the spot now, hope you can tell us more

Either by coincidence or just reuse/modify of an existing PCB, the ID resistors seem to be always next to one of the heatsink mounting holes.

Could be but I doubt it, when pulling high, the value starts at 8, not 9, both cards should in theory have the resistor in the same location, just a different value. It is more likely to do with the amount of RAM installed or some other configurable feature of the GPU.

Good spotting though, this resistor would now be one of my last to test if the others did not bear any results as it seems less likely to be the one for the device ID strapping.

[AlphaC]

gnif, you're awesome 

I have a few questions though.

Does this allow the modded card to get full performance of a Quadro in workstation applications? If yes, have you tried it on specviewperf11 and speccapc? (http://www.spec.org/benchmarks.html)

A GK104 GTX 680 4GB is the same as a K5000  ; GK107 GTX 650 2GB is the same as the K2000. The K4000 is iffy, the GK106 GTX 650 Ti /Boost has the same CUDA Cores though it doesn't have 3GB VRAM. I believe the Boost version is the direct correlation since it has 192-bit memory bus, but it's not out yet.

GeForce GTX 650 Ti    0x11C3 http://www.techpowerup.com/gpudb/2059/NVIDIA_GeForce_GTX_650_Ti_Boost.html
GeForce GTX 650 Ti    0x11C6http://www.techpowerup.com/gpudb/1188/NVIDIA_GeForce_GTX_650_Ti.html
vs
Quadro K4000    0x11FA http://www.techpowerup.com/gpudb/1841/NVIDIA_Quadro_K4000.html

GeForce GTX 650    0x0FC6 http://www.techpowerup.com/gpudb/894/NVIDIA_GeForce_GTX_650.html
vs
Quadro K2000    0x0FFE http://www.techpowerup.com/gpudb/1838/NVIDIA_Quadro_K2000.html
or
Quadro K2000D    0x0FF9 http://www.techpowerup.com/gpudb/2021/NVIDIA_Quadro_K2000D.html


From NVIDIA's Linux Drivers, http://www.nvidia.com/object/linux-display-amd64-310.40-driver.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/310.40/README/index.html ; http://us.download.nvidia.com/XFree86/Linux-x86_64/310.40/README/supportedchips.html

[gnif]

gnif, you're awesome 

I have a few questions though.

Does this allow the modded card to get full performance of a Quadro in workstation applications? If yes, have you tried it on specviewperf11 and speccapc? (http://www.spec.org/benchmarks.html)

A GK104 GTX 680 4GB is the same as a K5000  ; GK107 GTX 650 2GB is the same as the K2000. The K4000 is iffy, the GK106 GTX 650 Ti /Boost has the same CUDA Cores though it doesn't have 3GB VRAM. I believe the Boost version is the direct correlation since it has 192-bit memory bus, but it's not out yet.

GeForce GTX 650 Ti    0x11C3 http://www.techpowerup.com/gpudb/2059/NVIDIA_GeForce_GTX_650_Ti_Boost.html
GeForce GTX 650 Ti    0x11C6http://www.techpowerup.com/gpudb/1188/NVIDIA_GeForce_GTX_650_Ti.html
vs
Quadro K4000    0x11FA http://www.techpowerup.com/gpudb/1841/NVIDIA_Quadro_K4000.html

GeForce GTX 650    0x0FC6 http://www.techpowerup.com/gpudb/894/NVIDIA_GeForce_GTX_650.html
vs
Quadro K2000    0x0FFE http://www.techpowerup.com/gpudb/1838/NVIDIA_Quadro_K2000.html
or
Quadro K2000D    0x0FF9 http://www.techpowerup.com/gpudb/2021/NVIDIA_Quadro_K2000D.html


From NVIDIA's Linux Drivers, http://www.nvidia.com/object/linux-display-amd64-310.40-driver.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/310.40/README/index.html ; http://us.download.nvidia.com/XFree86/Linux-x86_64/310.40/README/supportedchips.html

Thanks .

At the moment since only GPU2 can be modded, benchmarking apps like these wont help any since you need to have the Quadro as the primary card. I do know however that in the GTX4xx series that the conversion using the BIOS mod to change the PCI Device ID did indeed give an enormous performance gain. Until I get either a GTX680 or another GTX690 to work on and find the straps on, there is little more I can do. As soon as the funds have been raised and I can find the straps for GPU1 I will do a heap of benchmarking and post all the results on these forums.

[olsenn]

I just discovered that this post is being covered on Tomshardware, haha: http://www.tomshardware.com/news/Nvidia-GTX-690-Quadro-K5000,21656.html

[amigo]

gnif,

Truly exemplary work!

Would you please post your card's brand and full model # as obviously not all cards are based on the reference design?

I'd also like other people to post their brand/model #s as well if they've successfully done a similar mod because those of us who might be looking at the mod need to know what exactly to look for.

Thanks.