How to remove Fermi and Kepler laser cuts to fully enable core. 680m success

I picked up a 680m for a backup card and to test out my theory on how to find and disable laser cuts. The operation was a success and now the card's core is now fully enabled and artifact free. vBIOS mods were unnecessary, but SVL7 modded vbios was used for extended overclocking limits. This mod method should work on any Fermi or Kepler card, but not later, as later cards do not have laser cuts.



The laser cuts are on the core packaging, not the core. They are not visible with the top coating present. Below is an image of my original deceased 680m with the top coating and capacitors removed. The laser cut which disables 1/8th of the core is boxed in green. Both sides need to be reconnected.



While I removed all of the coating, you just need to expose the green box. Even on the working card I exposed far more copper than necessary, as I didn't know which laser cuts mattered. You can see a cluster of 5 cuts to the upper left of the green box. Reconnecting these did nothing that I could observe. Maybe they are related to quadro features.

I pried off the steel chip guard around the core with a strait edge razor and heat gun. This is not necessary as the cut is very close to the core. You just need to peel off the plastic covering. Desktop cards will have a big IHS though, which will need to be removed.

I recommending wrapping sandpaper around a screwdriver to remove the insulator as shown.



Below is the laser cut exposed on the working card.


You can see I have solder on the card for where I tried bridging the cluster of 5 cuts. This was pointless. The 680m has an open memory cap pad just to the left of the core disabling laser cut. I may fill this cap in at some point. The cap I knocked off when removing the chip guard was on the opposite corner.



There is a tiny dab of solder bridging the laser cut. This took many tries. Typically I soldered the laser cut to the surrounding ground plane.

I haven't voltmodded the card yet, but I did upgrade capacitors and the memory inductor. I did run 3dm11 and took the #1 hwbot spot at 980 MHz despite @Johnksss somehow clocking his card up to 1130 MHz. I'd be lucky for 1080 MHz at 1.05V even with the cap mods which added about 20 MHz to core clocks.

https://hwbot.org/submission/4506549_khenglish_3dmark11___performance_geforce_gtx_680m_9536_marks

I wish to thank whoever made the old mobility 9800 core unlock guide many many years ago. This is basically the same concept.

 

Incredible work!

 

Thanks! If only I did this 8 years ago.

 

If only I had this mod back then.

Great Job!

 

This would have been a great mod for when the 480 and 780 ti were the top cards. Easier to win at LN2 benches when your core is bigger than everyone else's.

 

Very true!

 

Very cool stuff!

 

I added the 1 knocked off and 1 never installed memory caps on the core and reskinned it.



I am out of liquid metal (I thought I had a full spare tube...), which may be affecting clocks since the corners aren't properly covered. Overvolting memory from 1.35V to 1.52V only increased memory clocks from 5ghz to 5.4ghz, so memory overclocking on the 680m is still junk despite it having the same RAM chips, voltage, and buck converter setup as the 780m. With 1.35V good for 5ghz, 1.5V should be good for 5.6 to 6, so something is weird on the 680m.

https://www.3dmark.com/3dm11/13968623

Slight score improvement, but in no way worth the power. I'm going to dial the overvolt back to 1.4V after getting some overvolted core runs with added external cooling.

 

Awesome thread!
Wish this technique worked for newer GPUs or even CPUs.

Does anyone have a link to the old guide? Sounds like a fun read.

 

It may have been on the Dell community forums, which don't keep threads around very long. The mobility radeon 9800 could have its core config unlocked from 8/4 pixel/vertex pipes to 16/6, which was an enormous improvement. The laser cuts were much bigger and visible through the top insulator, so the mod is the same idea, just a bit easier. The magnifying glass I took a picture through is 10x magnification so after the 9800 the laser cuts got way smaller.

 

Thanks, found one article :
https://www.techpowerup.com/articles/100

So what about AMD? What was the last generation before they stopped performing laser cuts?

 

So I can get a GTX 480m and try to unlock 480 shaders over the 352?

 

That was actually my original plan. The 480m power phases are really weak and will likely blow up with core unlock and overclock though. Despite having 4 core phases it's only rated for 80 amps. That and Fermi can't drive a 4k60 display.

 

I don't know if they ever stopped. AMD typically sold more fully enabled cards though so less incentive to mod.

 

Damn That would have been the most fun to do this mod to. And 480m is quite cheap on ebay.

I only have 580m and 780m both already max cores.

Do they even laser cut? I know lots could be enabled within bios

 

HWInfo64 gives you ASIC quality if you look in video adapter, and I was curious if that would affect the percentage if you do this.

 

It's still be a fun mod. Just may wanna have some spare FETs in case they explode.

IDK if they laser cut. I do have a hardware disabled R9 390. I could check.

It won't. I do have a bit higher than average ASIC quality though. This 680m clocks significantly better than my original. I am tempted to try to reach 1260 Mhz on the core, as this would effectively be a 100% overclock with the core unlock included. Kepler seems to scale better with voltage over 1V than Maxwell.

 

I'm just gonna follow this thread
edit : I tried with my K4000M and it crash when the driver is installed, I'm going to increase a bit the voltage and see if it works

 

Can you include pictures? The K4000m should have more than 1 cut. I would think there's 4 total. I was surprised there were was only 1 trace visible on the 680m. I would have expected to see several intact traces.

Also I soldered the bridge multiple times, but also shorted it to the ground plane and had to remove and retry. Shorting may cause a driver crash even if the CUDA block is good. I could only see the short through 10x magnifying glass. It's too small to land a multimeter probe to check.

 

Well I just shorted the two pads you shorted too and that's it, I have a 675MX core which is the same as the K4000M and I have also a fully unlocked GTX770 core, I'm gonna remove capacitor then scrape everything to make picture
You won't believe me but I did it first time lol, I couldn't believe myself either, just a back and forth movement lol, it's not shorted to ground, it give me around 1700Ohm between the pad and ground


Edit : well it seems to works, only in optimus with a lot of artifacts but the core might be partly defective (from the factory) or the voltage might be too low (even though it's very high for this frequency), vbios is now the K4000M unlocked bios by svet from techinferno

Edit2 : memory are ok, no error from nvidia diag tool so that's core only artifacts

edit3 : My theory is that I partly unlocked the core, the core "thinks" all cuda are present so it tries to run them all but a part is still locked so there is some artifacts, furthermore I tried underclock of 50MHz and overclock of 300MHz (stable before) and it still artifact at the same rate.

edit4 : or maybe this jump is only used for debug/binning purpose at nvidia factory to bypass every "lock", because it feel strange that the difference between a fully unlocked GK104 and the 9th other core/memory configuration is this specific jump

 

Interesting. Yeah you had 5 disabled SMU blocks turn on with that one solder dot. It looks like it's kinda an override like you say. I did include a pic of my old 680m core. There are around 5 other laser cuts that I never touched. Maybe you have more?

 

Are you talking about these ? this is a picture of my 675MX core, I'm gonna sand paper my 770 core and mesure resistance between the 10 visible pads and compare it with my 675MX value, there might be something interesting here.

 

@TheQuentincc

Yeah the tight grouping of 5 I bridged to ground and there was no observable change. I did not try grounding the 5 single ones near the top of your picture.

The physical configuration looks the same as the 680m config.

 

GF110 GTX580

Edit : before scratching mf GF114 GTX570M, I can see 10 pads but I don't really want to scratch it

 

I think you may have killed that 570m. You sanded the die a bit, and you also took the edge off of the ground in the lower right, breaking a lot of vias to ground. I would always sand with just a sliver of sandpaper and made sure my finger did not run over the edge of the chip package.

 

this wasn't a 570M, it was a 580 GF110, it's one of my many 580's that does not work at all
So I mesured my GTX760 "debug" trace resistance and this is what I got (illustrated on your picture), I did the mod, now I'm gonna find it's heatsink and try it

edit : it does not work :/

 

I'm thinking that modding cores with 3/8 and 2/8 SMU blocks disabled are unlikely to have all fully functional disabled SMU blocks. Better luck with 680m and 670 with only 1 block disabled. 480 and 780 ti should be great candidates as there was no fully enabled card when those were released.

 

Sadly, my 765M also doesn't want to be unlocked, even if it already has vBIOS from 770M I have only artifacts.

 

Does the card not crash though? I have heard of unlocks result in many artifacts, but the card will render and never crash. Not ideal but kinda curious.

 

I will remove bridge today and check if card still works. Is there possibility, that vBIOS from MSI 770M 3GB (required to run this card in GT680) tries to address missing 1GB of memory after core unlocking?

 

Oh yes it would make sense that the memory controller for the missing memory turns on. Do you have a bunch of green lines horizontally across the screen? If so then that's what's happening.

 

It looks like that, different pattern after each power up, sometimes with erratic flashes. Pattern also changes after pressing power button, when laptop is turned on.

 

Ok I don't know what that is. Missing memory would give you a 75% correct image.

 

not really, sometime a gpu refuse to boot/display something with only one bad memory chip.
Your experience goes along my theory that it's a "debug" thing that fully enable the core, AFAIK bios does not unlock the core, a guy upgraded the core of his 1060 mxm to a 1070 + two more memory chip and it worked with the 1060 bios.

 

I've gotten that with AMD GPUs but not Nvidia, but maybe an address solder issue fully breaks it. With 1/4th of chips missing you'll get an image where 3/4th of the lines on the screen are normal and the others are a solid or wrong color. The BIOS POST doesn't even have Nvidia GPU memory initialized. There was someone with a 780 who removed all memory chips and it would still display POST artifact free.

 

I have k4000m as well, guess wouldn't be worth destroying if @TheQuentincc didn't get theirs working.

So how are you guys exposing those points, sanding the green layer around core?

 

sanding paper yes, P800 for me, cut to a small piece that I hold at the end of a cotton swabs and a bit of isopropyl alcohol, that's making a clean area and does not destroy the core.
Then for soldering I was having trouble with the 760 (my K4000M was so f*** lucky) so I used kapton tape all around the area, put flux on the area, a tiny bit of solder on the end of my iron, then solder those two pads, after that I checked the resistance between this to ground and it was 277k (if it's 0 that's bad, it should be 10+k)

edit : for now we don't know if this is unlocking a 1344 cuda/256b memory card to a 1536 cuda/256b memory only or not

 

@Khenglish @TheQuentincc Do either of you know how to modify nVidia vBIOS? I want to remove the UEFI (gop) portion of vBIOS from 9 series/Maxwell quadro based cards to test in older laptop.

I read something about header needs to be fixed or something? I am trying to learn vBIOS editing but it seems way beyond my comprehension to modify like that.

 

@ssj92 Nope, idk how to do that kind of mod.

After unlocking the 680m I did a lot of mods:

Without mods other than unlock, card was stable at stock voltage at 967 core 4700 ram. Kepler series cards clock the core in 13 MHz increments, with higher values always rounding down.

1. Switched out all 330µF caps for 470µF caps. Added missing 22µF ceramic cap for core. Switched 1µH memory inductor for .56µH inductor. Added 3 missing 2.2µF 19V caps.
Result: core stability increased to 980, memory to 5 GHz.

2. Added 60µF of ceramic caps directly under core. This helps Maxwell cards a lot.
Result: Only made 993 MHz less unstable.

3. Increased ram voltage to 1.51V.
Result:Ram stable clocks increased to 5400 MHz.

4. Added 7 missing 1µF RAM caps found on 780m and 880m, and 30 missing caps only found on the 980m. Piggy backed additional 470µF cap on ram caps. Added 2x 47µF 19V caps. 980m needs these to not lose ram overclock when core is overvolted. 680m has better 19V filtering though.
Result: No RAM clock improvement.

5. Piggy backed 7x 330µF caps on the 470µF core caps.
Result: Core stability increased to 993 MHz. 1006 MHz almost stable. When at 1.05V stability increased more from 1059 to 1085.

6. Reduced RAM voltage to 1.41V. Ram overvolt mostly just wasted power and card thermal throttles when running 4K at stock core voltage. Kepler is incredibly hot compared to Maxwell.
Result: RAM stable at 5175 MHz.

7. Added additional 70µF of ceramic caps directly under core.
Result: Core now stable at stock voltage at 1006 MHz.

Voltage filtering mods increased core clocks from 967 to 1006 MHz. 4% more speed for free. The only other mods I know of that work are adding wires to increase power efficiency between the core and inductors, but these mods are uuuuugglllyyy, are hard to do, and make the card more annoying to install. I consider mods done unless I want to replace the memory. I did find an ebay seller for 7GHz rated chips for $50 that doesn't appear to be a chip relabeling scam... Kepler IMC tops out around 7.5 GHz. idk if there is another issue other than ram quality which results in the low RAM clocks. Typically you expect 6GHz at 1.5V on older chips like these. Poor voltage scaling is a sign of poor ram chips. 980m boards run 6 and 7 GHz rated samsung chips at 1.35V both around 6GHz, but the 7GHz rated chips clock much higher at 1.5V.

 

You tempt me to buy a 680M and Epower it while watercooling it
Indeed Kepler card can consume a lot of power and be very hot, I barely manage to cool my 770 on watercooling while the vcore is at 1.45v and around 1450MHz, it consume about 450w in Firestrike GT1, thanksfully (but sadly too) I hit the OVP or the OCP on my card :/, I also believe it's core is pretty weak, it does 1350MHz at stock 1.21v but does not scale really well
Did you ever tried to mix memory chip ? like keeping the same brand but use 6 and 7GHz chip ? might need same chip revision ?
I think I read somewhere that kepler IMC is powered by the memory power phase, so increasing voltage should result in better clocks if everything is well cooled.

 

@TheQuentincc

For 28nm clock scaling beyond 1.1V is poor. My understanding is desktop stock volt Kepler cards tend to top out in the 1200s so 1350 at stock volts is excellent.

Mixing memory speed grades is fine. When switching memory chips on a card I do 2-4 at a time, then test the card. Doing all 16 and then having an issue is not fun.

Mixing memory brand is bad. Hynix and Samsung memory timings are highly incompatible and require major downclocking to work at all.

Mixing dies from the same manufacturer is questionable. I switched out 7GHz Samsung C dies for 8GHz E dies and lost clocks on a 980m board. I suspect the C timing table was not very friendly to the E die, but maybe the IMC needs to be designed for it, like Haswell designed DDR3 not clocking well on SB and IVB systems even when overvolted.

 

First generation of kepler seems to top out at 1250~1300 while the "second" generation goes to 1300~1350, out of 4 7xx kepler cards, I got at least 1300MHz at stock volt.
Yes you can get a very different timings between memory die, for exemple on ddr3 there is elpida hyper vs hynix mfr, one maxes out at 2000~2200 cas 6 while the other is more 2800+ cas 8 on air, to know which timings is used between die revision, look at amd bios, they list the timings, which shouldn't be so far off nvidia timings.

 

Yeah the 700 series does seem to clock a little higher than the 600 series. The 770 is gk104-425, instead of gk104-400, so maybe there was a mild die revision for higher clocks. Looking up 780m and 880m overclocking they do seem a bit better than 680m clocks on average. Now I am slightly inclined to replace the GPU core... I do have a substantially better than average first gen core though. I have seen 680m overclocking at stock voltage between 902 and 993 MHz. Mine was probably good for 980 prior to shader unlock. I never tested max overclock prior to unlock, but I would think more active core dropped it a 13 MHz bin. Not much more to be had by a 2nd gen core. More gains for less work to replace memory. Core is such a PITA to swap.

 

The hardest part for me when it come to swaping core is to reball the new chip...

 

Yeah that's annoying as hell. You'll get all but say 5 balls. Redo those 5 and have 2 left, then 1 left, then still 1 left... I got a fairly good method at this point so that reball only take me around half an hour to an hour. The biggest issue I have is random corners deciding not to settle properly leaving opens.

 

I just stumbled upon this thread by accident. Awesome work, as usual, Brother @Khenglish.

 

Absolutely incredible work, @Khenglish. I love this community (despite the sometimes overly hard anti-BGA sentiment here) because of how hard everyone here pushes their notebooks even if they're massively neutered, and this makes me want to push my 80 W RTX 5000 as much as I can.

Sucks to have a core count equal to the RTX 2080 Super, but have it perform like a RTX 2060...

 

You can always shunt mod it. You just need a soldering iron and you can order the shunts off mouser or digikey.

 

Yep, I've got shunt resistors arriving later this month. Hopefully I have a steady hand.

 

I'm usually lack finesse in this mod and I doubt if it can be applied to maxwell 2 980M BGA. My ASIC quality is 58% and OC's are pretty bad.
Nonetheless, as usual khenglish brings back performance crown from maxwell to Kepler.