I want to flash a GTX 1060 vbios onto a p106-90 6gb

[Transpanda]

I want to start off by saying: I know I can use the gpu with a patched driver and an igpu, I did that for the last half year (thank you dartraiden). I got an upgrade (1080 ti) and now want to do this for fun and to see if I can. Lastly I can afford to brick the card as it‘s €25 on ebay.

General infos:
From what I was able to find out, I might be able to use the modified nvflash version to ignore the certificate/board ID.
My card specifically is most likely out of a prebuilt mining rig from manli and can’t normally be bought individually.

Similarities and differences:
The vram amount, vram clock and the actual gpu/board including the additional 6-Pin power connector are the same as on a normal gtx 1060.
The differences are: no display out (missing components), pcie 1.1 x4 (might be upgradable?), lower TDP/lower core clock/less overclockable due to the low TDP and the important part 640 shader cores and 40 TMUs instead of 1280 and 80.

The vbios problem:
I was able to find some gtx 1060 and P106 BIOSes, but non of them specified the shader core or TMU amounts. So I’m wondering does a pascal generation bios include these amounts in the first place or does it just work with what the card has?
Because depending on that I have to either modify a pascal bios, which seems basically impossible, find a magical 1060 bios with the right amounts, go a completely different route or just use a normal 1060 vbios.

Any advice is appreciated

 

[StViolenceDay]

The 2 blocker differences are:

• different device id. GPU will just ignore VBIOS with the pcie deviceid different from its own. The device id (2variants diffing by 0x40 bit, but that doesn't matter here) is fused in the chip itself.
• The video output capability is not "just not soldered", the video-output-block inside the chip is fused-out (disabled) during gpu factory configuration process.

Initially all GP106 gpus are produced identical, and later are converted into a specific variant like 1060 or p106 via the "fusing/internal fuse burning" procedure, meant to be phisycally irreversable

 

[Transpanda]

GPU will just ignore VBIOS with the pcie deviceid different from its own

By that, do you mean it will refuse to save the vbios during the flashing process, or does it simply refuse to boot after, due to the mismatched vbios? Is there a way to avoid that like, tricking the gpu into reading it as being the correct id?

video-output-block inside the chip is fused-out

Thank you, that saves me from trying to replace the missing components, just to find out it doesn’t work haha

And just for my understanding: is the video-output-block specifically for the cards own ports? I’m asking because I was able get video out over an igpu, so im guessing it’s just for the ports.

Edit: and if I were able to flash it would the shader/TMU amount cause problems? I’m asking again, cause I found basically no documentation on this and I’m really curious.

 

[StViolenceDay]

Flashing is not the problem - even if tricks using nvflash would fail - it can be just done via cheap hardware SPI flasher configured/adjusted to 1.8level signaling and a SPI flash attachment clip. However the GPU would ignore the VBIOS with wrong device id during power-on/boot.

I mean internal video output block. "Copy the rendered image via PCIe bus back to main system" is working on P106 ming cards (with their original VBIOS) unless programmatically blocked in the driver.

Shader + TMU count does not cause any problems. TMU (and shaders inside them) are typically disabled via fusing, but physically diferent TMUs are disabled for different chips, often degraded ones. And VBIOS just doesn't touch them assuming that they are already correctly fused.

As an example - GTX1060 based on GP104-150 chip different samples has different memory channels (ROPs) disabled - note the absent VRAM chips in different positions in the attached photos (corresponding to KA/KB chip labelling)

But those GPUs ARE VBIOS-compatible: - any GP104-150 VBIOS would run on any GP104-150-K* chip. Same applies to TMU/shaders: vbios does not hardcode TMUs/shader count, it is setup during fusing stage.

And a note about PCIe x4 width - chances are this one may be not hardcoded in the fuses/vbios. Chances are that just adding the absent 12*2 = 24 of 200nF non-soldered-from-factory capactitors one the PCIe signal lanes may turn the GPU in PCIe x16 (still keeping the 1.1 version). I didn't tried this trick on P106, however it works on RTX3050s using RTX 3060s board. GA106-based RTX3050 upgradees from x8 to x16 just fine. No VBIOS change is required for such upgrade for 3050, si maybe with P106 too.

 

[Transpanda]

Flashing is not the problem - even if tricks using nvflash would fail - it can be just done via cheap hardware SPI flasher configured/adjusted to 1.8level signaling and a SPI flash attachment clip. However the GPU would ignore the VBIOS with wrong device id during power-on/boot.

I mean internal video output block. "Copy the rendered image via PCIe bus back to main system" is working on P106 ming cards (with their original VBIOS) unless programmatically blocked in the driver.

Shader + TMU count does not cause any problems. TMU (and shaders inside them) are typically disabled via fusing, but physically diferent TMUs are disabled for different chips, often degraded ones. And VBIOS just doesn't touch them assuming that they are already correctly fused.

As an example - GTX1060 based on GP104-150 chip different samples has different memory channels (ROPs) disabled - note the absent VRAM chips in different positions in the attached photos (corresponding to KA/KB chip labelling)

View attachment 368062
View attachment 368063
But those GPUs ARE VBIOS-compatible: - any GP104-150 VBIOS would run on any GP104-150-K* chip. Same applies to TMU/shaders: vbios does not hardcode TMUs/shader count, it is setup during fusing stage.

And a note about PCIe x4 width - chances are this one may be not hardcoded in the fuses/vbios. Chances are that just adding the absent 12*2 = 24 of 200nF non-soldered-from-factory capactitors one the PCIe signal lanes may turn the GPU in PCIe x16 (still keeping the 1.1 version). I didn't tried this trick on P106, however it works on RTX3050s using RTX 3060s board. GA106-based RTX3050 upgradees from x8 to x16 just fine. No VBIOS change is required for such upgrade for 3050, si maybe with P106 too.

Thanks I didn’t expect such a detailed answer. I didn’t know a good chunk of that.

So from what I’m getting. The bios flashing is possibly possible and if, theoretically, the gpu though the vbios had a matching device id. It would boot even though it’s missing half the cores?


About the lanes: I saw a video about someone fixing the hdmi port and the pcie lanes on an amd mining card. I was hoping at least one of them is possible, because besides the complicated setup you need to get video out the pcie speed was a big bottleneck in games. I’m guessing nvidia just wants to make extra sure we can’t game on the mining cards, by disabling the video output block, but I might try to reconnect the pcie lanes and see if that works.

On that note I’m guessing you need two capacitors for each lane due to the fact each lane is two connections, one from and one to the gpu?

Also I took pictures of the card in question:

 

[StViolenceDay]

It seems that your board has already all capacitors soldered, so it is limited to x4 in some different way, maybe "fused" too. Not sure if anything can be adjusted.

On that note I’m guessing you need two capacitors for each lane due to the fact each lane is two connections, one from and one to the gpu?

No, thats a high-bitrate signal, so its transmitted via differential pair - one physical lane is for negative signal, other for positive. "Diffirential pair signalling" is not GPU-specific you can read about it as a more general concept.

The other direction transmission corresponds to the lanes on another side of GPU. And since its another direction with the "GPU and motherboard chnaging their roles" - the capacitors are placed not on the GPU but on the motherboard, typically you can find 16 capacitor pairs near the PCIe slot.

 

[Macro Device]

And a note about PCIe x4 width - chances are this one may be not hardcoded in the fuses/vbios. Chances are that just adding the absent 12*2 = 24 of 200nF non-soldered-from-factory capactitors one the PCIe signal lanes may turn the GPU in PCIe x16 (still keeping the 1.1 version).

So its damage might be the cause for my GTX 970 running at x2 instead of supposed x16? Not that I care today, in the age of such GPUs being cheaper than dirt... But still.

To OP: there's virtually no way you make it a real 1060. PCI-e lanes are hard limited to 1.1 x4, you can't do anything about it other than planting a different die on the PCB. Also no way you enable video signal without planting a new GPU die. All in all, everything that could've made it a proper 1060 is either disabled or full-on laser wiped from the die.

 

[Transpanda]

It seems that your board has already all capacitors soldered, so it is limited to x4 in some different way, maybe "fused" too.

That‘s unfortunate, would be interesting to know why the resistors are there anyways.

there's virtually no way you make it a real 1060.

Right, I should probably be more specific about my end goal. So as stated earlier, I’ve previously used the card for gaming with a patched driver. I understand I won’t get video out and being stuck on pcie 1.1 x4 is unfortunate, but not a deal breaker.

My two goals are:
1. Unlock the watt limit to what a 1060 normally takes (120, 140 max) and unlock other settings like voltage.
2. Make it possible to use normal unpatched nvidia drivers and software.

In other words it’s not supposed to become a 1060, I just want it to be detected as one on any system and not get nerfed by the vbios.

 

[StViolenceDay]

So its damage might be the cause for my GTX 970 running at x2 instead of supposed x16? Not that I care today, in the age of such GPUs being cheaper than dirt... But still.

Its a quite common case that they are physically damaged, they are placed near the side of the GPU and may take an accident heat.
Since you have exactly x2 and not x4 - one of the two pairs that are "extending x2 to x4" can be broken. Non-physical damage (pure-electrical) on those capacitors is quite rare.
However having problems with solder balls under GPU or elecetrical problems in the GPU itself is also not-so-rare situation, this is much harder to investigate.

You can find those capacitors on yours GPU back, they are counted from the side having video ouputs (so the initial pair corresponds to x1 PCIe slot)
Skip 2 initial pairs (skip 2x2=4capacitors), and visually inspect the next 4 capacitors forming next 2 pairs. Touch them to ensure that they are not in "nearly-detached condition".

Unlock the watt limit to what a 1060 normally takes (120, 140 max)

As an a bit risky but high-chance-to-succed experiment - this can be done by "fooling" the current sensor by using "shunt mod" technique from overclockers community - stacking additinal resistors over the R005. The most risks are "is the power system powerfuk enough to handle such wattage without burning" and the same about the cooling system.

 

[Macro Device]

You can find those capacitors on yours GPU back, they are counted from the side having video ouputs (so the initial pair corresponds to x1 PCIe slot)
Skip 2 initial pairs (skip 2x2=4capacitors), and visually inspect the next 4 capacitors forming next 2 pairs. Touch them to ensure that they are not in "nearly-detached condition".

Okay, thanks for the heads-up.
Not currently at my disposal, will do a couple weeks later.

1. Unlock the watt limit to what a 1060 normally takes (120, 140 max) and unlock other settings like voltage.

As stated above, your only way of doing that is a shunt mod.

2. Make it possible to use normal unpatched nvidia drivers and software.

You need to summon engineers who made this tech and force them to revert software/hardware bans they uttered. Or reverse engineer it yourself but it'll take ages because no one besides said engineers knows where to dig. I wish I was joking. Pascal GPU unlock is super dead.

 

[Transpanda]

As multiple people suggested I’m probably going to try a shunt mod in order to get the core clock closer to what a normal 1060 gets, without getting unstable.

You need to summon engineers who made this tech

Do they want an offering? Or is it a prayer kind of thing?

On a serious note, thx for the heads up.

Pascal GPU unlock is super dead.

Kinda figured, that’s unfortunate

 

[Kogaraashi]

As multiple people suggested I’m probably going to try a shunt mod in order to get the core clock closer to what a normal 1060 gets, without getting unstable.

Do they want an offering? Or is it a prayer kind of thing?

On a serious note, thx for the heads up.


Kinda figured, that’s unfortunate

Good evening, sir. Could you please answer a few questions regarding the threads you've been commenting on? in your case have you tried to use the pci x16 in the 106 and 104 but in both cases you have failed, in my case I have the idea of buying a 102-100 10gb, but according to the threads that I have read some graphics cards are missing the capacitors to be able to have the power of x16 instead of x1, while in others despite placing the capacitors it still does not work, I saw from your comment that it was because the nvidia lock had not been broken, but today there are tools such as OMGVflash and NVflashk, and if now it is possible to use the pci x16 in a 102-100, and I wonder if you recommend me to buy the 102-100 for the price including shipping of 80 dollars, thank you very much for your answer.

 

[Transpanda]

Good evening, sir. Could you please answer a few questions regarding the threads you've been commenting on? in your case have you tried to use the pci x16 in the 106 and 104 but in both cases you have failed, in my case I have the idea of buying a 102-100 10gb, but according to the threads that I have read some graphics cards are missing the capacitors to be able to have the power of x16 instead of x1, while in others despite placing the capacitors it still does not work, I saw from your comment that it was because the nvidia lock had not been broken, but today there are tools such as OMGVflash and NVflashk, and if now it is possible to use the pci x16 in a 102-100, and I wonder if you recommend me to buy the 102-100 for the price including shipping of 80 dollars, thank you very much for your answer.

Hiii, (I’m a girl btw not a dude lol)
but concerning the p106, the capacitors for all lanes are present, but it still runs in pcie 1.1 x4 regardless. So its most likely fused right on the gpu die. In other words there’s absolutely no way to fix it aside from literally soldering on a new gpu to your pcb. Also flashing a different bios is possible, the issue is that the gpu will most likely refuse to boot. I‘m not sure how the pcie situation for the 102-100 is you should do some more research on that first.
Concerning your actual question if its for an experiment or secondary pc then sure go for it. But if its for your main rig it depends. If its difficult to get an equivalent gaming card in your area for a similar price and you’re fine with having to use all sorts of tricks for it to work then yes get it, its a good card for a good price. On the other hand if you can get a regular card for just a little more money and/or you are bothered by having to work on it, having a slow pcie connection, needing a second gpu or igpu, possibly having limited framerate, possibly lacking some features like video encoding for vr or remote play then maybe stick to the normal cards.

also in case you’re interested my p106 is currently attached to a thinclient over nvme also i did the shuntmod and overclocked it +222mhz core +500mhz ram (tested 100% stable)

 

[t3rbin]

I too wanted to try this bullshit many years ago, but then I lost the desire because the p106-90L is very weak anyway, more similar to a 1050 Ti than a 1060. I thought they had already found a way to fix PCI-E 16x and that the only thing left to was enable Gen 3 support. Modified drivers aren't a problem since Pascal is about to go out of support. However, I don't know much about this GPU and it would be disappointing to waste time with the possibility of achieving little or nothing. if it costs next to nothing there must be a reason.