Its most likely one of the ram chips is either faulty or has an issue with the lead free solder balls.

You can read here for some quick info on various failure modes. But some of it depends on what era and what type card you're diagnosing. For example Voodoo cards with a separate texture mapping chip and dedicated texture RAM.

http://www.playtool.com/pages/artifacts/artifacts.html

If only the GPU is BGA you can use an oscilloscope on the RAM chips to see wether there are plausible signals present on all legs to differentiate between contact issues and broken RAM chips.
If the artifacts are caused by a bad solder joint pressing the chip (with heatsink) against the PCB will often cause them to change or disappear. This usually won't be the case with broken RAM chips. That's a cruder approach but who has got an oscilloscope anyway? 😉
Lastly sometimes a faulty RAM chip will work again if cooled down with ice spray which can help in identifying the individual culprit.
Apart from that it's really not easy to differentiate between a contact issue and broken components on BGA chips where you can't just measure for continuity.

This is a very broad question, diagnosis is different for cards of different eras and architectures.

It also depends on where it comes from. If it was tossed around in a recycling center for weeks then it could be physical damage as well.
SMD parts are easily knocked off, pins bent, traces scratched up and even RAM chips break quite easily.
IMO the majority of "modern" GPU failures are due to a cooler malfunction. These chips don't have THERMTRIP# so they just cook themselves to death.

What can it mean if the card shows artifacts in its default settings (in 3d mode), but the artifacts disappear when I reduce the memory frequency by 10MHz? Is this permanent memory corruption? Or maybe a problem with its power supply?

dulu wrote on 2021-07-07, 17:37:

What can it mean if the card shows artifacts in its default settings (in 3d mode), but the artifacts disappear when I reduce the memory frequency by 10MHz? Is this permanent memory corruption? Or maybe a problem with its power supply?

Both is possible. What card are we talking about? Can you make good pictures of front and back and post them here?

dulu wrote on 2021-07-07, 17:37:

What can it mean if the card shows artifacts in its default settings (in 3d mode), but the artifacts disappear when I reduce the memory frequency by 10MHz? Is this permanent memory corruption? Or maybe a problem with its power supply?

VRam most likely...😉

Depends on the types of artifacts as well..

wiretap wrote on 2021-07-07, 18:15:

Depends on the types of artifacts as well..

Pics would be nice. A picture is worth a thousand words. It is true...😉

I say in advance that this gpu is not the reason I started the topic. This is the Medion Radeon 9800XL

The problem is that it's hard to "catch" these artifacts in the photo. Both via camera and screenshot. IRL, you can see many, many more.

I tend to think if it will display 3d than its not the gpu. If you see things like vertical strips it can be bad ram.

Interesting insights in this thread, thanks!
Here's a somewhat related idea that came to me a while back. I thought you guys might find it amusing 😁

Being retro fans, most of you have probably heard of Visual 6502. That's a transistor-level simulation of the famous 80s chip (done in Javascript no less!).
What about doing something similar for old graphics cards? Or indeed, other old parts like memory chips?
Obviously it's not very feasible with the current tech, simply because for the most part you'd need to re that info from (very) high-resolution photographs. That's not the only constrait, but I guess that would be the main one.
To put things in perspective, the 6502 was first manufactured with a 8 micron process.

But maybe possible some day? Perhaps having open source hardware could help.
The logical next step from that would then be to teach an AI to classify all the different kinds of visual artifacts and link them to whatever part of the hw was falling apart.