Hello,
I'm having a problem repairing the AGP version of HD3850. The card in Windows crashes with driver installed, but doesn't display any artifacts. At this point I have replaced all VRAM chips, no change whatsoever. Tests fail only at tested 10MB and more. I wonder if memory controller in GPU is damaged.
In direct-mem-test.py and dmgg.py I get these reports.
I was thinking about disabling banks with VREF, but there are no artifacts, so that won't be helpful. Is there anything else I can do to repair it?

very sorry, but this is a very old video card. it is better to buy a used working one than to restore this one ((

That's the point, I'm trying to repair the most powerful AGP card, as these are quite rare and expensive :)

OK, a quick update: The only "place" where there is an error is when I test 10 or 11 MB of VRAM (31 and 511 error count respectively). There are no errors when testing 9MB or less and there are 511 errors when testing 11MB or more, up to 512MB of VRAM.

This pointed me to a broken solder joint either under VRAM (not possible because of new ICs installed, each one reballed with leaded solder) or under the core itself. Reflow didn't do anything - same errors in same places (error bits).

I seriously doubt that the cores memory controller is damaged in such a way that only *probably* only one bit is damaged, and chances of the VRAM IC in the same place being damaged are also pretty low.

There might be damage to some traces under the core or under VRAM, which I didn't see, hence why it throws this errors, but that's also extremely unlikely.

If noone has any other idea, then I deem this GPU unfixable.

Did you have display output throw the same card during test? If yes - those 31 and 511 errors are just false-positives caused by cursor blinking/other visual output activity.
Try with outputting the picture from integrated video or an extra GPU in PCI slot (however I'm not sure how such old GPU behaves in such case, but worth a try).

So, I suppose the actual test result is "There is no any basic errors with memory" detected by those .py tools.

I've been testing it on a 939Dual-SATA2, and using a second GPU for display purposes didn't even cross my mind.

Surely enough, after testing this card with a Nvidia PCI-E GPU as primary dmgg doesn't show any errors.

However, the GPU still crashes after loading drivers on my test setup with a GA-7N400 motherboard.

BUT it doesn't crash on the 939Dual mobo, with/without drivers, with/without load the GPU is always rock-stable.

Considering these circumstances I wonder if there even was a problem to begin with. Well, you live, you do, you fail, you learn.
Anyway, thanks for pointing me in the right direction, I got so fixated on the card itself that I didn't even consider that other things may fail.