• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

I broke it! | RTX 3090 Water block install gone wrong

Soldato
Joined
19 Oct 2004
Posts
4,201
Location
London
Having just seen the ebay listing, it's not actually as bad as I expected from some of the descriptions. The trace damage to the PCB doesn't look that extreme and most of the damaged components can be sourced from RS Components for a few pounds. I've seen worse examples on other forums that came back to life. That said, people bidding £700 for it are being a bit optimistic - you could in theory get a 3080 for that price and having something 15% slower that you 100% know is working.

It definitely is worth picking a slow day when it comes to GPU modding. I had to re-solder my 3090 FE about fives times until I got the results I wanted and then all the hassle of the paraphernalia along the way - new soldering iron, better flux, resistors without a sloping shelf on the shunt, magnifying glass and thermal camera (hotspots). Always stressful, even when you're used to it.

what did you do to it? Shunt mod? Or something more extreme?
 
Associate
Joined
9 Apr 2017
Posts
188
Location
Eve Online
what did you do to it? Shunt mod? Or something more extreme?

Just shunt modding, and varying the resistors on the different shunts to get the best results (especially the pcie one). Also tried doing the additional 'small' 5mohm stacks along with the 6 main ones (very fiddly and not really necessary in the end). The process takes longer because I wanted to do some of them one by one to see which shunt equates to pin 1, pin 2, pwr_src (linked to pcie in the end), etc since there were some things I couldn't get from a multimeter. And because I didn't have my waterblock at the time, re-assembling the damn FE cooler with Fujipoly extreme pads each time.. it's a pain, and the FE pcb is very dense making for tricky soldering work. Everything needs 20 minutes of kapton tape time before starting to make sure solder doesn't contact onto the nearly touching components.

But there's no practical power limit now, and I can run the same clocks for top 30 HOF scores on Cyperpunk all evening, no issues yet. It can pull 650-700w just for the card but Cyperpunk is running well on Psycho RT settings, so there's some return for the work. Also used it to capture some 12k gaming.
 
Soldato
Joined
19 Oct 2004
Posts
4,201
Location
London
Just shunt modding, and varying the resistors on the different shunts to get the best results (especially the pcie one). Also tried doing the additional 'small' 5mohm stacks along with the 6 main ones (very fiddly and not really necessary in the end). The process takes longer because I wanted to do some of them one by one to see which shunt equates to pin 1, pin 2, pwr_src (linked to pcie in the end), etc since there were some things I couldn't get from a multimeter. And because I didn't have my waterblock at the time, re-assembling the damn FE cooler with Fujipoly extreme pads each time.. it's a pain, and the FE pcb is very dense making for tricky soldering work. Everything needs 20 minutes of kapton tape time before starting to make sure solder doesn't contact onto the nearly touching components.

But there's no practical power limit now, and I can run the same clocks for top 30 HOF scores on Cyperpunk all evening, no issues yet. It can pull 650-700w just for the card but Cyperpunk is running well on Psycho RT settings, so there's some return for the work. Also used it to capture some 12k gaming.

impressive.
 
Soldato
Joined
24 Sep 2013
Posts
2,890
Location
Exmouth, Devon
Just shunt modding, and varying the resistors on the different shunts to get the best results (especially the pcie one). Also tried doing the additional 'small' 5mohm stacks along with the 6 main ones (very fiddly and not really necessary in the end). The process takes longer because I wanted to do some of them one by one to see which shunt equates to pin 1, pin 2, pwr_src (linked to pcie in the end), etc since there were some things I couldn't get from a multimeter. And because I didn't have my waterblock at the time, re-assembling the damn FE cooler with Fujipoly extreme pads each time.. it's a pain, and the FE pcb is very dense making for tricky soldering work. Everything needs 20 minutes of kapton tape time before starting to make sure solder doesn't contact onto the nearly touching components.

But there's no practical power limit now, and I can run the same clocks for top 30 HOF scores on Cyperpunk all evening, no issues yet. It can pull 650-700w just for the card but Cyperpunk is running well on Psycho RT settings, so there's some return for the work. Also used it to capture some 12k gaming.


650-700W! Flippin' eck!
 

P.B

P.B

Soldato
Joined
2 Apr 2015
Posts
2,530
Just shunt modding, and varying the resistors on the different shunts to get the best results (especially the pcie one). Also tried doing the additional 'small' 5mohm stacks along with the 6 main ones (very fiddly and not really necessary in the end). The process takes longer because I wanted to do some of them one by one to see which shunt equates to pin 1, pin 2, pwr_src (linked to pcie in the end), etc since there were some things I couldn't get from a multimeter. And because I didn't have my waterblock at the time, re-assembling the damn FE cooler with Fujipoly extreme pads each time.. it's a pain, and the FE pcb is very dense making for tricky soldering work. Everything needs 20 minutes of kapton tape time before starting to make sure solder doesn't contact onto the nearly touching components.

But there's no practical power limit now, and I can run the same clocks for top 30 HOF scores on Cyperpunk all evening, no issues yet. It can pull 650-700w just for the card but Cyperpunk is running well on Psycho RT settings, so there's some return for the work. Also used it to capture some 12k gaming.

Damn that gives me hope for my fe if I ever wanted to shunt mod it
 
Associate
Joined
9 Apr 2017
Posts
188
Location
Eve Online
Damn that gives me hope for my fe if I ever wanted to shunt mod it

Yeah, the FE can go very well; it's trading blows with Kingpins and modded Strix's, #24 on Timespy Extreme and #7 on Wild Life, hopefully get up there in Port Royale too, which isn't bad for the same 24/7 clocks I game on. Having said that, for once I actually feel bad taking the stock cooler off - Nvidia have really upped their game there. It worked really well for an air cooler and felt absolutely premium. I tend to put function over form but this time around I thought most of the AIB efforts looked incredibly gaudy and silly.

650-700W! Flippin' eck!

It's definitely not efficient but there is something satisfying about seeing the power numbers under load. PSUs were a problem - with the last mods it somehow trips the OCP on my Asus Thor 1200w. But it works flawlessly an EVGA 1600 T2, which I think it due to the analogue controller (the Asus is digital).

For a long time I was stuck on 580w due to hitting some kind of MVDCC/chip power limit and I'm still not 100% sure which of the mods finally overcame that.

Can't believe it sold for £1550!, Hopefully it's not a bot!

Going back on topic, that's just incredible if a card with that kind of damage sold for nearly rrp. There probably are people who can repair it (imo from the pictures, the damage is limited to the kind of components you can buy on RS/digikey, etc.) but we're still talking days of someone's time to get everything running smoothly again.
 
Associate
OP
Joined
14 Jul 2018
Posts
21
Highest bidder pulled out unsurprisingly! Sent second chance offer to second highest bidder, and thats at £1500, soo will see if they get back to me, but I'm not holding my breath!
 
Associate
Joined
7 Apr 2006
Posts
292
Location
USA
Just shunt modding, and varying the resistors on the different shunts to get the best results (especially the pcie one). Also tried doing the additional 'small' 5mohm stacks along with the 6 main ones (very fiddly and not really necessary in the end). The process takes longer because I wanted to do some of them one by one to see which shunt equates to pin 1, pin 2, pwr_src (linked to pcie in the end), etc since there were some things I couldn't get from a multimeter. And because I didn't have my waterblock at the time, re-assembling the damn FE cooler with Fujipoly extreme pads each time.. it's a pain, and the FE pcb is very dense making for tricky soldering work. Everything needs 20 minutes of kapton tape time before starting to make sure solder doesn't contact onto the nearly touching components.

But there's no practical power limit now, and I can run the same clocks for top 30 HOF scores on Cyperpunk all evening, no issues yet. It can pull 650-700w just for the card but Cyperpunk is running well on Psycho RT settings, so there's some return for the work. Also used it to capture some 12k gaming.

I sent a PM,
How did you bypass the hardwired internal power limit that triggers at ~600W, that isn't shuntable?
This rail "seems" to be reported in HWinfo64 as the bottom-most GPU Core (NVVDD) Output Power (sum), and I already know this rail completely ignores shunt mods.
I also know what the sum of this rail is. It gets its sum from the bottom-most GPU Core (NVVDD) Output Power, and SRAM Output Power in HWinfo64.

SRAM output power isn't a problem as that's linked to SRAM Input power, which seems to respond to one of the shuntable rails. It's the other rail that seems to ignore everything, I've seen that in excees of 360W reported, when "total board power" (shunted) is only reporting 300W (before the 1.9x multiplier correction for 570W to the card from the wall).

For a long time I was stuck on 580w due to hitting some kind of MVDCC/chip power limit and I'm still not 100% sure which of the mods finally overcame that.
*BINGO!*. The rail causing you problems is that output rail I mentioned. It's not MVDDC nor Chip Power Limit. I found all the rails where chip power is summed from and none of them are hitting a limit.

Are you sure it wasn't one of the "tiny" 005 shunts ? Because you said modding those wasn't necessary. But perhaps it is...? Can you elaborate on what complete mods you did? Because that's the issue limiting a BUNCH of cards...FE cards, Palit cards, MSI cards, all shunted....it's that strange power rail ....

If you can precisely determine exactly which mod "fixed" the throttle, this will be the most important shunt mod find of 2021. Because that is limiting ALL shunted cards. Even the STRIX is affected.

*edit again*:

GPU Chip Power = GPU Misc1 Input Power + GPU Misc3 Input Power + GPU Core NVVDD1 Input Power (sum).

GPU Core NVVDD2 Input Power(sum) = GPU Chip Power + GPU Sram Input Power (sum)

GPU Core NVVDD Output Power (sum) = GPU Core NVVDD Output Power (the very bottom one in Hwinfo64) + GPU SRAM Output Power.

The "hidden" rail causing throttle SEEMS to be GPU Core NVVDD Output Power, which when added to SRAM Output Power, feeds GPU Core NVVDD Output Power (sum).
 
Last edited:
Soldato
Joined
3 Oct 2007
Posts
12,085
Location
London, UK
I sent a PM,
How did you bypass the hardwired internal power limit that triggers at ~600W, that isn't shuntable?
This rail "seems" to be reported in HWinfo64 as the bottom-most GPU Core (NVVDD) Output Power (sum), and I already know this rail completely ignores shunt mods.
I also know what the sum of this rail is. It gets its sum from the bottom-most GPU Core (NVVDD) Output Power, and SRAM Output Power in HWinfo64.

SRAM output power isn't a problem as that's linked to SRAM Input power, which seems to respond to one of the shuntable rails. It's the other rail that seems to ignore everything, I've seen that in excees of 360W reported, when "total board power" (shunted) is only reporting 300W (before the 1.9x multiplier correction for 570W to the card from the wall).


*BINGO!*. The rail causing you problems is that output rail I mentioned. It's not MVDDC nor Chip Power Limit. I found all the rails where chip power is summed from and none of them are hitting a limit.

Are you sure it wasn't one of the "tiny" 005 shunts ? Because you said modding those wasn't necessary. But perhaps it is...? Can you elaborate on what complete mods you did? Because that's the issue limiting a BUNCH of cards...FE cards, Palit cards, MSI cards, all shunted....it's that strange power rail ....

If you can precisely determine exactly which mod "fixed" the throttle, this will be the most important shunt mod find of 2021. Because that is limiting ALL shunted cards. Even the STRIX is affected.

*edit again*:

GPU Chip Power = GPU Misc1 Input Power + GPU Misc3 Input Power + GPU Core NVVDD1 Input Power (sum).

GPU Core NVVDD2 Input Power(sum) = GPU Chip Power + GPU Sram Input Power (sum)

GPU Core NVVDD Output Power (sum) = GPU Core NVVDD Output Power (the very bottom one in Hwinfo64) + GPU SRAM Output Power.

The "hidden" rail causing throttle SEEMS to be GPU Core NVVDD Output Power, which when added to SRAM Output Power, feeds GPU Core NVVDD Output Power (sum).

This is the place to ask. Also they have the KingPin bios which is 1000w I believe. That will only work with 3x 8 pin 3090 cards though

https://www.overclock.net/threads/official-nvidia-rtx-3090-owners-club.1753930/
 
Associate
Joined
7 Apr 2006
Posts
292
Location
USA
This is the place to ask. Also they have the KingPin bios which is 1000w I believe. That will only work with 3x 8 pin 3090 cards though

https://www.overclock.net/threads/official-nvidia-rtx-3090-owners-club.1753930/

No one on OCN knows. Multiple people have run into that limit on various shunted cards, all triggering between 520-580W.

Maldoror is the only person who successfully bypassed that internal power limit without the Kingpin bios (which cannot be flashed on FE due to protection), so I have no choice but to ask here and hope he replies.
 
Soldato
Joined
24 Aug 2013
Posts
4,549
Location
Lincolnshire
I don't mind you laughing at me, honest!

Basically, never attempt to do something like this while having a blazing row with your girlfriend!

I ordered this ASUS STRIX OC RTX 3090 in September 2020, it arrived Saturday 16th January 2021. Had all the stuff ready to go on the card, EK water block and back plate.

I've done this a few times before, and never had an issue. But this time its gone all wrong. All because I was to heavy handed in removing the existing backplate (due to girlfriend row! My fault, not hers!), which caught two capacitors on the board and ripped them off. The attempt to reattach them to the board obviously didn't go well!

I test fitted the card in my other gaming PC, this particular 3090 was destined for my sim racing rig to run my new HP Reverb G2 (which is amazing BTW!). But as soon as the PC was turned on, there was much white smoke, followed by a lovey smell of burning electronics, you know exactly the smell I am talking about don't you!

So yeah, I failed massively with this card!

Question is, what can I do with it? No warranty, as number one I've evidently removed the warranty sticker, and number two, this was my fault, not NVIDIA's!

I obviously don't have the skill to repair it myself. :D
So do I stick it on eBay and hope for the best?
Use it as a door stop, or an expensive paperweight?
Find someone who has a far better chance of repairing it?
Forget NVIDIA and wait for the new AMD card (more waiting :rolleyes:)

Shame (I never saw the original photo) but they can be very difficult to solder for the inexperienced used. Not just because of the size but because they have 3-400w power planes so will soak the heat from a soldering iron really easily, making them difficult to solder.

Ideally need preheating, soldering and then fully measuring with a meter afterwards. Any shorts will show then you can start again if required.

Next time just be super careful, I always take hours when building up wc’d cards. Or if in doubt purchase from ocuk. They do cards equipped with Waterblocks that they build up themselves. And offer warranty.
 
Last edited:
Back
Top Bottom