Graphics Card Necromancy - EVGA 980 Ti Classified

[RazorWind]

I want to start this thread by first thanking the [H] member who sent this card to me at his own cost, "no strings," as he put it. It sounded like he (or she? I'm not sure) wanted to remain anonymous, so I won't tag him here, but I hope you'll all join me in thanking him for sending it to me. I'm sure he could have sold this on ebay for a few bucks, but instead, he chose to send it to me so we can all learn something. I know I speak for all of us when I say I'm grateful, but I hope you'll do so as well. Anyway, because this was so generously donated by an [H] member, I'm working on the thread first, and I'll upload the video to Youtube later. I suspect most of us here prefer to read, rather than watch, anyway.

The card in question is an EVGA GeForce 980 Ti "Classified," a product of the uniquely famous Ilya 'TiN' Tsemenko and Vince 'Kingpin' Lucido duo at EVGA. This was one of the best iterations of the 980 Ti that you could get at the time, and as we'll see later, its very well made, with a lot of features aimed at overclockers (arguably the "real" hardware enthusiasts) that gamers shouldn't be that eager to pay extra for. It's similar to the Kingpin edition ("KPE" as Kingpin likes to call it), and you can see they're closely related if you look at them side by side. The KPE really goes ham with the overclocking features, whereas they appear to have been a bit more reasonable with this card. One thing I think is really neat about this card is that even now, when Nvidia has made it extra difficult, EVGA still really leans in to the extreme overclocker crowd, and not only employs TiN and Kingpin for the purpose of making special cards like this, but allows them to be accessible to the community in ways that none of the other board vendors do. MSI and Galax used to have similar product lines (called "Lightning" and "Hall of Fame" respectively), but it's really only EVGA that's still obviously putting in this much effort. Furthermore, TiN, at least, is a shockingly accessible person. He apparently owns and operates Xdevs.com, where he makes similar content to mine, and will sometimes respond if you reach out to him and ask a question.

Here's the card as it arrived. It clearly got used, as it's got some dust in the fans, and the heatsink fasteners show some evidence of having been removed at some point, but it doesn't appear to have been abused, or used for GPU mining in an outdoor shed or anything. Frankly, this is exactly what I would expect this card to look like at this point in its life, having been owned by the sort of person it's intended for. It even still has the peely-film on it.


The BIOS switch was in the liquid nitrogen mode when it arrived. Giggity. 😎



Those headers are among the things that make this card special. The one on the left is what EVGA calls the "Probe-it" header, which consists of a series of voltage read points. With the exception of the 5V rail, every one of the voltage rails on the card is present there. There is a wiring harness that plugs into this; I'm not sure if it comes with the card, but we don't happen to have it. For our purposes, we can just probe the pins directly, so that's no biggie. For competitive overclocking, you'd want that wiring harness.

The other header is a connector for what EVGA calls an "EVBot," which is an external controller gizmo that allows real time control of the card's voltage rails. EVGA doesn't seem to make them anymore, but it has some advantages over using software that runs on the host system, and it's a cool feature that they include on this card.

A visual inspection of the card reveals a bit of physical damage here, near the PCI-E finger, where we can see we're missing a couple of 0402-size SMD capacitors. I don't know if this damage happened before or after the card was shipped to me. It seems likely that it may have happened before, as I've been pretty careful handling it, but I can't rule out the possibility that I did this. Fortunately, this is a pretty common thing, and should be relatively easy to fix. It might also help explain why the card doesn't work. Those little caps actually are critical, as they're in series with the PCI-E data lines. I've mused about this before, but I wonder if this sort of damage is one of the reasons that those metal shrouds are becoming common on newer cards. I bet the cost of including that is less than an RMA.

 

[RazorWind]

Now, what the previous owner told me is that the card will work in Windows, but crashes as soon as you try to run a game. That suggests some sort of logic issue. In my own testing, it doesn't work at all, perhaps as a result of the damage to the PCI-E interface. I suspect there's more going on, but we'll see.

Our first diagnostic step is to check resistance to ground on each of the board's voltage rails.

Here's the first 12V rail. That's 55Mohms, which is unusually high. We'll need to remember this and investigate later. It's not a short to ground, though, so we're not going to risk a fire if we plug this thing in and power it up.


Here's the other. This one looks OK, at about 2.5K.


Availing myself of the handy-dandy Probe-it header, we'll check the other rails. I also sanity checked these against the actual logic, but I really love the Probe-it header, so that's what I'm showing here. I wish every card had this.

VCore looks good. Remember that we're looking for single digits here.


Memory... looks wacky. Oddly, if you look at the docs for the EVGA EPower V, they claim that the resistance on this rail is actually supposed to be 15 ohms, which is also quite low, so it's possible that this is within spec, but I expect this is probably related to whatever the original problem was. On a "normal" 980 Ti, like the MSI Golden Edition we looked at before, this would be more like 75 ohms.


Vpll looks OK as far as I can tell.


3.3V looks... Kinda low. This could be normal, though.


The last one is the 12V rails again. I gather this is wired up to the far 8 pin connector, as it shows a normal-ish resistance, and not the millions we saw on the near one.

 

[RazorWind]

Alright, sorry for the radio silence, folks.

The next step was to plug the card into the test bench and power it up. We get no picture on the screen, but we also get no sparks or fire, so at least there's that.

We've got 12V on both connectors...


I availed myself of the super handy Probe-It connector for most of the rail measurements. I also confirmed these readings at the real components closer to the logic.

We've got a sane, if not exactly high, VCore. This is ~930mV or .93V. Spec is something like .7 to 1.2, so the card should be able to run on this.


Vmem is present, but a bit below spec. We've got 1.409V here; spec is 1.5. This may be the reason our card doesn't run. On most GPU designs, there's an enable signal that the VRMs are supposed to provide to the GPU to indicate that they're ready, which they won't provide if the voltage is low.


VPll looks OK. This is 1.05 volts


Edit: 3.3V. Looks spot-on to me.


And our 12V reading on the Probe-It connector. I later figured out that this is supplied by the PCI-E connector, not the 8 pins.


Next step is to remove the heatsink and have a look at the front of the card...

 

[oldmanbal]

you had me at "The BIOS switch was in the liquid nitrogen mode "

 

[RazorWind]

you had me at "The BIOS switch was in the liquid nitrogen mode "

Indeed. This card has clearly seen some shit.

Anyway, I got the heatsink off.





It's somewhat less impressive in 2020 than it was in 2015, but this is still a freaking awesome card. That's a fourteen phase VRM, fitted with Infineon/IR DirectFET 6721s on the high side and 6725s on the low side. IR's DirectFETs are pretty interesting because the metal surface you see on the outside of them is actually the drain terminal, meaning that while you need to be careful that you don't put a metal heatsink directly on top of them, they're much easier to cool than FETs that have a ceramic or epoxy casing. Each phase is capable of providing about 50W, meaning you could conceivably provide 700W to the core, as long as you had enough cooling. For most users, this just means that it should be able to run in the nice friendly part of the efficiency curve, but it also means that you wouldn't be hurting for power with exotic cooling, such as dry ice or liquid nitrogen.

It took some doing to clean the card once I got the heatsink off. It was shockingly difficult to photograph, but it appears that in its past life, this card had liquid metal applied to it as a TIM, which was apparently not removed very thoroughly before a regular gray aluminum oxide thermal grease was applied on top of it. I started cleaning it, thinking this was just normal Arctic Silver or something...


But I was surprised when some globs of very shiny, very heavy material appeared from the pool of alcohol and grease residue around the die.


It took forever to get it off - normal solvents don't seem to work, but I finally got the bulk of it.


I admit, I'd never used liquid metal before, although I think I did encounter something like it once on one of my 295X2s. I assume it must be pretty amazing as a TIM. Given how much of a pain in the neck it was to clean off the card, I can't imagine anyone would use it if it didn't at least work really well.

Some of the liquid metal appears to have gotten on the thermal pads at some point, which suggests it escaped from the GPU die, and may have gotten into BGA connections under the GPU or memory ICs.


I don't see any under the corresponding memory ICs, but it's possible I cleaned it off, but it damaged the solder balls, or that it's further in there where I can't see it. That might explain our low resistance on the memory power rail, but is purely conjecture at this point.

 

[RazorWind]

Lacking anything specific and obviously wrong with the rest of the card, I decided to at least fix the broken off 0402s on the PCI-E interface.

I cannibalized four caps from this 290X, which we couldn't save previously. It already bravely donated one of its VCore bootstrap caps to save another 290.


And here they are on the patient card. Quarter shown for scale.


I then tested the card again, but we still get no post. I think this pretty much confirms that the real reason it doesn't work is related to our low memory voltage.


Unfortunately, a voltage rail that runs but is just low is one of the hardest problems to diagnose that we might encounter, at least in this case because we don't have a datasheet for the controller. We could, with great difficulty, figure out what on it is connected to what, but without the datasheet, there's almost no way we'd be able to troubleshoot it.
I'll have to give some thought to what to do next. We're kind of stuck without datasheets for anything.

 

[RazorWind]

So, after getting my butt kicked by Solan's MSI card all day last sunday, I pulled this card out and had another look at it. We've got a couple of suspiciously low resistance readings on a couple of its rails, and I'm struggling to figure out if they're just quirks of this card (the Kingpin version has similarly low resistance on those rails by design) or if they're indicative of a short.

I know it's a long shot, but would anyone happen to have a working one of these 980 Ti Classified cards? If so, would you be willing to take some measurements on it for me?

Specifically, I need to know what the resistance on a functional card is from the memory rail and the 3.3V rail. Both can be measured without disassembling the card at the Probe-it connector.

 

[balintbalint]

So, after getting my butt kicked by Solan's MSI card all day last sunday, I pulled this card out and had another look at it. We've got a couple of suspiciously low resistance readings on a couple of its rails, and I'm struggling to figure out if they're just quirks of this card (the Kingpin version has similarly low resistance on those rails by design) or if they're indicative of a short.

I know it's a long shot, but would anyone happen to have a working one of these 980 Ti Classified cards? If so, would you be willing to take some measurements on it for me?

Specifically, I need to know what the resistance on a functional card is from the memory rail and the 3.3V rail. Both can be measured without disassembling the card at the Probe-it connector.

I don't know, if it helps, but I have two "Gigabyte GTX 980 Ti G1 Gaming" with different memory chips, and i measured both of them. One comes with Hynix memory, and has 62 ohms on the memory rail, the other one with Samsung chips, has only 15 ohms on the memory rail. So, i think, 10 ohms could be possible.

 

[RazorWind]

I don't know, if it helps, but I have two "Gigabyte GTX 980 Ti G1 Gaming" with different memory chips, and i measured both of them. One comes with Hynix memory, and has 62 ohms on the memory rail, the other one with Samsung chips, has only 15 ohms on the memory rail. So, i think, 10 ohms could be possible.

That's actually very helpful. It would still be useful to get a measurement from a known-good Classified, but this is definitely a useful data point.

Unfortunately, what it seems to suggest is that one or more of the memory ICs on my Classified board is bad (or more likely, got some liquid metal under it or something).

 

[Krazy925]

That's actually very helpful. It would still be useful to get a measurement from a known-good Classified, but this is definitely a useful data point.

Unfortunately, what it seems to suggest is that one or more of the memory ICs on my Classified board is bad (or more likely, got some liquid metal under it or something).

I actually have an EVGA classified gaming card with a factory AIO unit somewhere in my closet.

It might actually be loaned out though now that I think about it. I can check though, with explicit instructions since EE isn't my gig.

 

[RazorWind]

I actually have an EVGA classified gaming card with a factory AIO unit somewhere in my closet.

It might actually be loaned out though now that I think about it. I can check though, with explicit instructions since EE isn't my gig.

Yes, thanks, that would be super helpful!

Here's what you do:

Grab a multimeter and set it to resistance (Ohms) mode. That's the omega symbol. Most default to auto ranging, which is fine.


Our first measurement is the memory power rail. This can be measured in many places, but with the card assembled, the easiest place on this card is likely the probe-it connector. I've marked the pin you want in this photo. With the card face down on the table, it's the third pin from the left.


To take a measurement, put the red probe on the pin or solder pads, and the black probe on something that's connected to ground, the metal casing around the dvi connector. With handheld meters, the reading will swim a bit. That's OK. I believe the expected value should be around 15 ohms. Whatever you get, post it here.

Our next measurement is resistance through the 3.3V rail. For this, the best place to measure is the PCI-E connector. Put your red probe on any of the pads I've marked here, and the black probe on any ground, such as the outside of the DVI connector. On most cards, you'd expect to see between 200 and 1000 ohms. Whatever reading you get, post it here.

 

[Mode13]

Hey razorwind, I'm really new to this / just started studying EE so yeah... I don't really quite understand at all how you know what proper resistance values are, but ..I'll learn eventually, provided I can find the right reading materials..

Anyway, I came into an EVGA 780Ti classified that's not working (VGA not found post code on the PC) and I'm messing around with it.. It was working when it was packed and mailed to me (gift from friend), but arrived toast.

I'm testing the spots you asked for previously in case they matter between the cards, the PCB design doesn't seem too drastically different based on your pictures, 14x 6725 & 6721 mosfets making up the core VRM.. Though the 780Ti one does have SK Hynix memory. I thought maybe my readings may help you, and maybe you would have some input for me... Hopefully I don't make your situation worse by posting potential bad values from a card that no longer works.

Probeit pin 1 (GPU) -- 4.1 ohms
Probeit pin 3 (Mem) -- 84.5 ohms
Probeit pin 5 (PLL) -- 311.1 ohms

3.3V rail on the pcie connector finger -- 342 ohms

Interestingly, the +12V pins on the 8 pin power connector like you measured before are too difficult for me to get any meaningful reading on my Fluke 112.. the bottom right pin on the lower PCIE connector (same orientation as your OP) is pretty solid in the 2.3 M ohm region. The same pin on the top connector seems to want to float around too much for my meter to make up it's mind.. I'll get 3 to 5 k ohm, lift the red probe, put it back down, and then it's 3-5 M ohms and swimming.. ? Using the DVI connector shielding for ground..

Now, on mine, when I power the system on, 4 of the 5 LEDs light up, the one that doesn't light up is PLL.

With the system powered on I measured voltages at those same Probe it pins and got 0.9V on the core, 1.6V on the hynix memory, and 0.05V on the PLL. Close enough to 12V and 3.3V on pins 9 and 7 though I don't remember the exact values.

Issue on my end if you care .. I have no idea where to proceed with this PLL voltage. According to EVGA --
"PLL - This indicates if the voltage for auxillary components such as fans, LEDs, etc. is good"
Not sure that would keep the card from working and the fans are spinning just fine.. I thought PLLs were clock generators that would be use to set the operating frequency? Anyway.. No idea how to proceed and can't find anything at all online, short of a shot of a GTX 1080 with a "1V PLL voltage regulator" just below the memory.. There is a 1R0 ?inductor I believe? there that spits out roughly the same resistance as the PLL pin.. I kind of wish the issue was with core or memory voltage, at least those things are documented a bit more. I'm flying blind. No damage to the card at all, though there are a few questionable pads on the back of the card.. I'm waiting for some high res shots of a working card to compare with.

 

[RazorWind]

Hey razorwind, I'm really new to this / just started studying EE so yeah... I don't really quite understand at all how you know what proper resistance values are, but ..I'll learn eventually, provided I can find the right reading materials..

Anyway, I came into an EVGA 780Ti classified that's not working (VGA not found post code on the PC) and I'm messing around with it.. It was working when it was packed and mailed to me (gift from friend), but arrived toast.

I'm testing the spots you asked for previously in case they matter between the cards, the PCB design doesn't seem too drastically different based on your pictures, 14x 6725 & 6721 mosfets making up the core VRM.. Though the 780Ti one does have SK Hynix memory. I thought maybe my readings may help you, and maybe you would have some input for me... Hopefully I don't make your situation worse by posting potential bad values from a card that no longer works.

Probeit pin 1 (GPU) -- 4.1 ohms
Probeit pin 3 (Mem) -- 84.5 ohms
Probeit pin 5 (PLL) -- 311.1 ohms

3.3V rail on the pcie connector finger -- 342 ohms

Interestingly, the +12V pins on the 8 pin power connector like you measured before are too difficult for me to get any meaningful reading on my Fluke 112.. the bottom right pin on the lower PCIE connector (same orientation as your OP) is pretty solid in the 2.3 M ohm region. The same pin on the top connector seems to want to float around too much for my meter to make up it's mind.. I'll get 3 to 5 k ohm, lift the red probe, put it back down, and then it's 3-5 M ohms and swimming.. ? Using the DVI connector shielding for ground..

Now, on mine, when I power the system on, 4 of the 5 LEDs light up, the one that doesn't light up is PLL.

With the system powered on I measured voltages at those same Probe it pins and got 0.9V on the core, 1.6V on the hynix memory, and 0.05V on the PLL. Close enough to 12V and 3.3V on pins 9 and 7 though I don't remember the exact values.

Issue on my end if you care .. I have no idea where to proceed with this PLL voltage. According to EVGA --
"PLL - This indicates if the voltage for auxillary components such as fans, LEDs, etc. is good"
Not sure that would keep the card from working and the fans are spinning just fine.. I thought PLLs were clock generators that would be use to set the operating frequency? Anyway.. No idea how to proceed and can't find anything at all online, short of a shot of a GTX 1080 with a "1V PLL voltage regulator" just below the memory.. There is a 1R0 ?inductor I believe? there that spits out roughly the same resistance as the PLL pin.. I kind of wish the issue was with core or memory voltage, at least those things are documented a bit more. I'm flying blind. No damage to the card at all, though there are a few questionable pads on the back of the card.. I'm waiting for some high res shots of a working card to compare with.

The resistance values are something you eventually get a feel for if you look at enough cards, as it's pretty uniform. Single digits for the core, double digits for the memory, 200 - 400 on the 3.3V, thousands to millions for the 12V rails, etc. Smaller GPUs will have higher core resistance than big ones, but the memory rail resistance should be similar.

I think the next step I'd take with your card is to test for voltage at the inductor for the PLL rail, and if you don't have voltage there, troubleshoot that VRM, which is likely much simpler than the core or memory. On my card, the inductor for this is the small 1uH one between the memory chips and PCI-E connector latch.

If you do have voltage (should be ~1.0V) there, then you need to troubleshoot whatever it connects to, which is likely to be quite difficult without a schematic.

 

[Mode13]

Thanks for the reply

I think I see what you mean.. the overall impedence is in some relatively predictable range..

have to get a test bench going, too annoying to mess with that part of the card with (what I believe) is the PLL vrm while it's inside of a case

Worse yet, I did a closer inspection of the card and counted AT LEAST 10 surface mount components knocked off of the back (top side) of the card.. we've concluded the customs must have opened the card up and played Frisbee with it.. I've just reached the point in my studies elsewhere where ordering SMD components and a hot air gun / nozzle / giant magnifying glass will be a necessity.. so I'll probably not come back to this for a few months until I have a bit of experience ruining lesser projects. The same guy that gave me this card may send me yet another 780Ti to use as a donor card so I can get the values off of the resistors or capacitors this one is missing.. I can't tell by eyeballing which size is which yet, but some of these suckers are TINY...

Maybe in a few months I'll have my own necromancy thread