How to identify the defects on this card?

[Nikan]

Hello folks,

This is my first post on this forum so please excuse me if I'm doing anything wrong.
BTW I'm studying embedded systems so I know a bit about electronics...
I have a few bench power supply's, an oscilloscop and a few multimeters, so please don't get too far with the measurements... I don't have a spectrometer etc. but if it's needed I can probably get my hands on it...
I also have an ERSA Soldering Iron, Rework station (IR and hot air) and a Reflow oven (IR) If needed

So I've got a broken GPU out of ebay, and now I want to try and fix it for educational purposes.
So the first thing I've done was taking it apart and look around for bad solder joints.
After that I was looking for magic liquid on the board but I coudn't any!
So the next step was to connect the board to my bench power supply and check for a short circuit.
Well the fans didn't work until 11V which tells me, that under-voltage-lockout circuitry is working fine!
The next thing I've done after I didn't notice any exsessive current consumtion was connecting the card to a mother board.
The card jumped right on, the only problem were some artefacts... I've got horizontallines while booting.
After I got into Windows the artefacts dissapeared, but I wasn't able to change the resolution. It was locked on 800x600 HDMI.
So I took the card apart again and found a brown RAM-block well it was light yellow (pretty hard to identify) I don't know if I didt that, or if it was just me not noticing it...
So the next step was to take the RAM chip off and connect the card back to the motherboard and as I've expected... the artefacts were still there.
After that I checked the GPU and It was showing only about 2GB of RAM Which is weird. I mean aren't the RAM chips allocated separately?
If yes, shouldn't the GPU detect bad RAM and shunt it?? If that is true I can be sure that something else is wrong too!

BTW GPU-Z is detecting the gpu as a GTX 980 so that is a good sign..., but ROPs unknown, texture fillrate unknown, bus width unknown, bandwidth unknown

Does that mean that the GPU is half dead or is it in some kind of a safe mode due to RAM error???

What should I do next?

I apprecate any Answer!
Thanks in advance! 

[Rasz]

1 take a picture, what does "brown RAM-block well it was light yellow " mean?
2 source replacement chip

[kripton2035]

heat the chip in a sort of reflow process.
this fix can then works for some months until the defects will start again.

[Muxr]

After that I checked the GPU and It was showing only about 2GB of RAM Which is weird. I mean aren't the RAM chips allocated separately?
If yes, shouldn't the GPU detect bad RAM and shunt it?? If that is true I can be sure that something else is wrong too!

Each GDDR5 memory chip typically has a 32 bit wide memory bus. But the gtx 980 has a 256-bit wide memory bus. So that should tell you that the memory is laid out in parallel fashion. Let me guess: 8 memory chips x 32-bit (per chip) = 256-bit (gtx980 spec)?

Typically a memory chip isn't going to fail with a visible physical defect. Usually on a GPU all memory chips are powered from a single VRM, so if the VRM failed short for instance all memory chips would likely fry (not just one).

[Rasz]

heat the chip in a sort of reflow process.
this fix can then works for some months until the defects will start again.

"sort of reflow" your balls dude :/

1 its not a fix
2 it "works" for hours to days to maybe weeks
3 it only works on defective GPUs, gtx980 is not one of them

[Nikan]

Well this is the chip I've taken off.
I hope you can see the same I'm seeing

https://picload.org/view/riiwrcli/dsc_0187.jpg.html
https://picload.org/view/riiwricl/dsc_0188.jpg.html
https://picload.org/view/riiwrpgr/dsc_0189.jpg.html

I think this chip died by thermal stress...

How can we check if the GPU is working properly? A µC for example either works or it doesn't... well sometimes some pins dont work, but if they are used in the circuit, the circuit won't work and we know either the mcu or the circuit around is broke. And we can just measure the outpu of the mcu to see if the mcu or something else is dead. how can we achive that with the GPU???

[tooki]

You know this forum supports image attachments? Then you don't have to use a third-party image service that displays the image like this on an iPad (hardly an obscure device). (For real, what are they thinking, putting ads onto -- and ONLY onto -- the single content area of the page??)

[Rasz]

That is rather curious indeed.
Chip in question is only $6 already pre balled on ebay, on the other hand they all look tired. Have you measured ram supply rail?

Edit: nope, its the 6000Mbps version, so from the GTX780 card, you need 7000 one. Unless you underclock your card

[Nikan]

I know, it is stupid as hell!
Yea true, but I took this pic with my nikon and those pics are almost 8MB and the maximum supported on this site is only 1Mb...
I'm sorry for that!

[Nikan]

Yea I know I've seen them... The only problem here is that I dont know how many died...
GPU-Z is only reading about 2GB I dont know if half of the RAM allocation is disabled for safety or if 4 of these chips are dead...
I would like to know that before I purchase just one and wait 3 months till it arrives only to see that more of them are dead...

[coromonadalix]

are you willing to go that far for a gtx780 ??  even for educational purposes,  you have to know when to quit ?? 

[Muxr]

imgur.com doesn't have annoying popups and it will resize the image for appropriate size automatically.

[Nikan]

1. It's a 980 but I actually don't care what it is...
2. Yes I'm cause every thing you fix makes you smarter!

[tooki]

I know, it is stupid as hell!
Yea true, but I took this pic with my nikon and those pics are almost 8MB and the maximum supported on this site is only 1Mb...
I'm sorry for that!

Everyone here has to deal with this, since even a crappy smartphone produces pictures bigger than 1MB. So resize them real quick and then upload, it's not that time consuming. 

(I much prefer when people upload pix here than use an external service, because those services may or may not be around, causing images to go missing in the future, as we see here all the time on threads a few years old. That doesn't happen when they're uploaded here.)

[Nikan]

Here are the pics

[Nikan]

2

[Nikan]

3

[Rasz]

measure c141 with card running to see if Vram is present/correct
980 might have so much cache on the die it might even try to pretend to work with no external ram = stupid results in GPU-Z
from my experience its closer to 1 month for chinese package to arrive

[Nikan]

Alright, I'll measure that cap when I'm home...
I hope there will be something wrong otherwise...

[EvilGeniusSkis]

You know this forum supports image attachments? Then you don't have to use a third-party image service that displays the image like this on an iPad (hardly an obscure device). (For real, what are they thinking, putting ads onto -- and ONLY onto -- the single content area of the page??)

Or use the

Code: [Select]

 [img] [/img]

tags to embed the image.