GPU Upgrade via BGA Core Swap Success - MR 9600

[timohour]

next is a GF108 core to a GF117 core on an NVS 5200m, which is a full node die shrink of the same core.

Is there any possibility that a upgrade from a GF108 (96 cores 40nm) to GK107 (192 or 384 cores 28nm) would be successful?

They seem to have identical socket.

would it destroy my motherboard if I tried?

[Khenglish]

Is there any possibility that a upgrade from a GF108 (96 cores 40nm) to GK107 (192 or 384 cores 28nm) would be successful?

They seem to have identical socket.

would it destroy my motherboard if I tried?

I don't think so since they expect a 128 or 192 bit memory interface.

- - - Updated - - -

I did not notice this thread earlier.

In brief. After my experience with the GK104 chip exchange in damaged K5000M for a GTX680 desktop chip (1203 A2), it seems to me that the number of total GPU active cores does not depend on the same chip, which in GK104 is almost identical for the desktop and for the mobile GPUs. It seems to me that it's only bios-power section depends.

[ATTACH=CONFIG]15363[/ATTACH]

Very interesting. I know that quadro features can be enabled by changing resistor values on the board. So it looks like resistors might be responsible for disabling shader blocks as well.

I had ruled this out before since this would mean that you cannot disable shader blocks for yield reasons, since I highly doubt pcb manufacturers are customizing the card for each individual core.

Now I really wish I compared resistors on the one 780m I fixed for someone vs my dead 680m.

[timohour]

I don't think so since they expect a 128 or 192 bit memory interface.

I may not understand your answer but if this list is accurate it seems almost all Nvidia cards that use the gf108 chip use 128 bit memory interface except for the NVS5200M which uses 64 bit.

The same goes for the GK107 chip which uses 128 bit memory interface on almost all cards except from Quadro 410 and Quadro K500M (64 bit).

example from an gt755M

Also, there is an mobile card that uses one of those two chips (GEFORCE GT640M LE) either gf108 or gk107.

And finally their socket seems identical (at least from the pictures I could find). The only problem that I could find would be the vbios of the card since it is intergrated in my laptop's bios.

Correct me if I am wrong but there is no gk107 chip implementation that uses 192 bit. (unless there is wrong info on the list...)

[Khenglish]

I may not understand your answer but if this list is accurate it seems almost all Nvidia cards that use the gf108 chip use 128 bit memory interface except for the NVS5200M which uses 64 bit.

The same goes for the GK107 chip which uses 128 bit memory interface on almost all cards except from Quadro 410 and Quadro K500M (64 bit).

Also, there is an mobile card that uses one of those two chips (GEFORCE GT640M LE) either gf108 or gk107.

And finally their socket seems identical (at least from the pictures I could find). The only problem that I could find would be the vbios of the card since it is intergrated in my laptop's bios.

Correct me if I am wrong but there is no gk107 chip implementation that uses 192 bit. (unless there is wrong info on the list...)

I feel that jumping to kepler would be too much. It would definitely need a custom vBIOS. That's why I was only looking at fermi chips. They are also the only chips to have a 64-bit memory interface. I was thinking that the memory was cut down on package, but it may be on board since Clyde's desktop 680 core swap had reduced core count.

As for 192-bit I was referencing the GF106 used on the 555m.

[timohour]

I feel that jumping to kepler would be too much. It would definitely need a custom vBIOS.

As for 192-bit I was referencing the GF106 used on the 555m.

I get this.

There is a dead E6430 motherboard on eBay for £36 on eBay. If things change and I can complete a PayPal transaction I may grab it and try. Worst case senario I would try the GF117 mod.

[Khenglish]

I get this.

There is a dead E6430 motherboard on eBay for £36 on eBay. If things change and I can complete a PayPal transaction I may grab it and try.

The reason I haven't tried it is because I'd end up killing all the round electrolytic caps on the motherboard. A heat gun cannot provide an even enough temperature without overheating part of the die package, and my oven would mean heating the whole board, which would kill those caps. The only possibility is using leaded solder to allow heating to lower temperatures, but then the bonds would be very weak as the leftover unleaded solder would not melt and bond well to the leaded solder.

Unless you have a real BGA station or are willing to replace all the caps I would not try it.

[Clyde]

I don't think so since they expect a 128 or 192 bit memory interface.

- - - Updated - - -

Very interesting. I know that quadro features can be enabled by changing resistor values on the board. So it looks like resistors might be responsible for disabling shader blocks as well.

I had ruled this out before since this would mean that you cannot disable shader blocks for yield reasons, since I highly doubt pcb manufacturers are customizing the card for each individual core.

Now I really wish I compared resistors on the one 780m I fixed for someone vs my dead 680m.

It's not so easy. Firstly, PCB of K5000M differs from GTX680M presence of the U505 and its surroundings. Secondly, I compared the resistors with my K5000M QS. I found nothing. I'm old and tired so I could have missed something, but in my opinion, that PCB is responsible for the number of active cores. At least in K5000M case.

[Khenglish]

It's not so easy. Firstly, PCB of K5000M differs from GTX680M presence of the U505 and its surroundings. Secondly, I compared the resistors with my K5000M QS. I found nothing. I'm old and tired so I could have missed something, but in my opinion, that PCB is responsible for the number of active cores. At least in K5000M case.

I just looked at some pictures and the 680m is lacking the R49 resistor just below the core on the left on the top side of the card, while the K5000m has it.

Did you check resistor values? for quadro vs geforce the same resistors are present, just values differ.

[MOVED] Hacking NVidia Cards into their Professional Counterparts - Page 1

[Clyde]

R49 can be just a U505 surroundings or ID resistor like you said.

I compared the values of resistors between broken K5KM and my K5KM QS. Except the number of active cores there were no differences between both cards which confirms bios, nv inspector and GPUZ. Conclusion, the card has a damaged or different power section traces on PCB. Anyway, in K5000M resistors id have no relevance to the number of active cores.

Unfortunately I do not have GTX680M at home to compare.

[Khenglish]

R49 can be just a U505 surroundings or ID resistor like you said.

I compared the values of resistors between broken K5KM and my K5KM QS. Except the number of active cores there were no differences between both cards which confirms bios, nv inspector and GPUZ. Conclusion, the card has a damaged or different power section traces on PCB. Anyway, in K5000M resistors id have no relevance to the number of active cores.

Unfortunately I do not have GTX680M at home to compare.

Interesting. I can't think of why it would be running fewer cores than a stock card.

[Clyde]

I'm not sure. I think it's K4000M board. BTW R49 in K5KM is a jumper 0 ohm.

[Clyde]

I find something. I cant find first byte resistor of device ID but I think that I find second. This resistor had 30K ohm (11BD) in my broken Quadro which corresponds to K4000M, 25K ohm (11BC) in K5000M QS and 45K ohm in GTX780M where it should have 40K (119F)???

I have not 5K resistor for try to change K5000M device ID for K5100M (11B8).

[Clyde]

I finally found a 5K resistor (can be 2X10K in parallel).

1. After soldering 5K Ohm resistor instead 25K Ohm on my K5000M QS with K5000M vbios, Windows automatically installs graphics drivers and GPU has 1,536 cores. However, the card is visible everywhere as the standard VGA.

2. After soldering vbios chip with Clevo K5100M vbios, the drivers are installed only after INF mod. The K5000M QS is now K5100M with 1344 cores.

Without drivers

With drivers

[@t0mX]

Love this Thread! So many things can be tried, only problem is the required skills and tools to be able to unsolder and resolder correctly the BGA chips on our boards! I don't think that's an "everybody can do this" kind of skill, like "bake yourself some BGAs instead of eggs" ! anyway, good luck on the tests guys, awesome work!

[Khenglish]

I finally found a 5K resistor (can be 2X10K in parallel).

1. After soldering 5K Ohm resistor instead 25K Ohm on my K5000M QS with K5000M vbios, Windows automatically installs graphics drivers and GPU has 1,536 cores. However, the card is visible everywhere as the standard VGA.

2. After soldering vbios chip with Clevo K5100M vbios, the drivers are installed only after INF mod. The K5000M QS is now K5100M with 1344 cores.

Without drivers

[ATTACH=CONFIG]15521[/ATTACH] [ATTACH=CONFIG]15527[/ATTACH]

With drivers

[ATTACH=CONFIG]15523[/ATTACH] [ATTACH=CONFIG]15522[/ATTACH]

Wow, it sounds like we almost have this. It seems like the Nvidia driver never installed when you had the K5000M vBIOS, and that you only got the basic WDDM video driver.

With the K5100M vBIOS the fact that it reverted to 1344 is nuts. To me this is indicating that the driver can still somehow see that the PCB is a K5000M, or that maybe something is messed up in the driver since you previously were running a K5000M. You did see a gain shader count when changing the resistor value, so we're on to something. Performance is in line with 1344 shaders I'm guessing?

I wonder what shifting to a Desktop K5000 ID would do. This would likely be necessary if this mod were to work on a 980m 1536 to 2048 shader mod. The 970m/980m ID resistors are in the same region as the kepler cards, but in a different configuration. I think the ID resistors are the two closest to the lower-left corner of the pcb. They have a 3rd solder point indicating that they can be shifted between VCC and GND. The 970m is shifted from the 980m which is necessary due to the different ID.

I'm thinking that the resistor table listed in the quadro thread might not be 100% correct. Ex. maybe 0xF isn't 40k Ohm, but anything 40k Ohm or higher.

I am now trying to find my own card to play with. Maybe the core swap was unnecessary?

[Khenglish]

I'm also concerned that the K5100M vBIOS may not be correct for the card, and maybe that is causing a problem. For example, the 680m cannot run a 780m vBIOS.

I would try the K5000M vBIOS again and try to get the driver to install properly.

[Clyde]

First time, I was not able to calmly verify the intermediate stage, namely K5000M vbios and 5K ohm resistor. It was not me soldered, I leave such details skilled in the art. I'm already soldering back EEPROM with K5000M vbios but the POST or something is stuck. Tried everything and I can not repeat this scenario. I do not understand why. I saw 1,536 cores in GPUZ after driver auto-installation. GPUZ shows 1536 cores also with K5100M vbios before the drivers installing so there is no certainty that it does not do auto-refilling error?

When it comes to performance of K5100M 1344 cores was only possible execution of Specviewperf11.

What is normal for K5000M 1344 cores stock.

Unigine Heaven barely walked. 3DMark11 barely walked and finally crashing. K5100M 1344 cores is back K5000M QS.

Resistor table could by OK (if R >= 39.5K).

From the beginning I knew that such a swap may not work well (redesigned PCB, 4 vs 8 GB of GDDR5, the lack of a second resistor ID). After early experiments with exchange of the GPU chip from desktop to mobile, I was almost certain that nVidia uses the same chips for almost all Kepler's GPU. One, what I've learned from Clevo, that it is not the hardware but just the license matters. Therefore, I wanted to confirm that the change of the number of cores running depend on resistors ID and vbios, not of the chip itself. Unfortunately, the experiment did not fully succeed because I am not convinced that GPUZ after soldering 5K resistor really seen 1536 cores whether, only reading K5100M device ID from resistor (11B8 instead of 11BC), automatically complement the data, but at least my K5000M QS was for a while a K5100M and it is still great.

[Clyde]

Recently I got hold of GTX680M, which I was able to convert in K5000M.

In front:

1. You need to unsolder the 5K (R42) resistor marked in red.

2. You need to solder the 25K (R40) resistor and 0K jumper (R49) in places marked in green.

In back:

1. You need to unsolder two 100K resistors marked in red (on the left) and 35K (on the right).

2. You need to solder 35K resistor in places marked in green.

Resistor marked in white (on the right) is a GDDR5 manufacturer ID resistor where 45K is Samsung and 35K is Hynix assigned.

3DMark11 stock + 135/1000MHz

Only BIOS has some problem yet with the proper card recognition.

Have fun.

[H658tu]

Very nice !

How did you find out the required resistors and values? You went from '11A0' to '11BC', right? So ... is mobile different from desktop or does each core have a specific R->Hex table?

[Clyde]

Very nice !

How did you find out the required resistors and values? You went from '11A0' to '11BC', right? So ... is mobile different from desktop or does each core have a specific R->Hex table?

Simply, using a multimeter and logic. Seriously, I made a resistors mirror image of my both K5000M and GTX680M pcbs. Yes But I not proved anything yet because the card still works just like the GTX680M. Quadro drivers normally recognize both systems and start installation but only the Ge Force drivers are installed, and in SPECviewperf11 card has performance of GTX too.

Unnecessarily I changed also one of the resistors because GPUZ shows that the card has Hynix GDDR5. In fact it has GDDR5 from Samsung.

Edit: Resistor marked in white on the back of de card (on the right of the picture above) is a GDDR5 manufacturer ID resistor where 45K is Samsung and 35 is Hynix assigned.

[H658tu]

Cheater !

Seriously, not so good news then. Flashed Quadro vbios?

For the memory; hopefully there's a non-Samsung 680M somewhere ... or, actually, any identical ***M MXM with different brands. That'd show soon enough where these are identified.

[R01991]

Hello Khenglish. I try to find anything but i failed. Could you tell us if you got any success with 7970m bga swap ?

Best regards,

Sebastian.