This post is going to be a compilation about a 3 AGP GPUs, which mostly didn’t go as planned.
Case 1: Radeon HD3850
This GPU is the most powerful card You can realistically get. I’ve seen around the internet AGP HD3870, but it’s literally factory overclocked 3850.
This card worked fine untill the driver has been installed. Easy fix – dead VRAM. Surely enough dmgg.py shows VRAM errors.
A quick BGA process later…
We have no change.
After that I changed every VRAM IC, to no success. Noticing that these errors occur only at 10 and 11 MB of tested memory I decided that there has to be a BGA issue on one bit somewhere and subsequently reflowed the core.
BUT THEN:
Thermocouple got soldered into a trace!
Nothing I couldn’t handle. Man, I have to get myself a microscope, microsoldering leaves a toll on my eyesight.
Is it the prettiest?
Not really.
Does it work?
N. O.
Wait, what? Even reflowing the core didn’t help?
No.
At this point someone from Vlab told me to try it as a secondary GPU.
So I used a 939Dual-SATA2. Surely enough, 0 errors, even the driver loads correctly. No artifacts, no problems at all.
…
GPU likely was fine, only the stupid Gigabyte board was too exclusive for it.
Case 2: Radeon 9800 XXL
These cards are notorious for dying – VRAM, core, You name it. Everything fails. In this case too.
This thing is a trip. Even though dmgg says that there’s only 1 IC bad, this REALLY doesn’t look like one. It looks like 2 or 3 are bad. Foreshadowing?
Also, as You can see, armed with experience I tested this card as a secondary GPU.
My chips are a bit newer, but otherwise are similar enough.
Reballing in progress…
As per dmgg I changed the top one.
Nope… Not fixed.
Bottom one?
Even worse.
Great. Now 2 chips throw errors. These might have been on the verge of dying and heating them up to about 120 degrees just made them quit.
Because these VRAM ICs are so unreliable I think I’ll be better off replacing each and every chip… or stealing the core to put on another PCB, with VRAM from a different manufacturer.
Case 3: Radeon X1950 PRO AGP
This GPU arrived as “untested”. Fortunately it worked, so in this case only a refurbrishing process was needed.
VRAM thermopads are still squishy and wet, so I decided not to replace them and only change the thermal paste on the core.
A PCI-E logo on an AGP card, interesting…
These VT1165SF mosfets needed a different pad, but it was a weird thickness, about 1.2mm. I think 1mm pad would be sufficient, but to be sure I used a Upsiren thermal putty.
As You can see, the GPU works just fine. There is one thing though…
Addition: Bridge cooling
The AGP to PCIE bridges get very hot, but due to space constraints they didn’t get any heatsinks from the factory. I’m going to fix that.
My old GPU, the HD3650. I swear, it’s a coincidence that all GPUs are ATI 🙂
A bit of thermal putty and thermally conductive glue…
… and a radiator on top of that. This glue is not conductive, so it doesn’t really matter if it leaks out. Due to the difference of height between the die and radiator, as well as thickness added by the putty I like to add a lot of glue – this way it’s glued for sure.
Almost the same goes for the X1950.
Glue, because there is no insulator it’s also on the die.
… and the radiator on top.
Before You ask, the 3850 also got a heatsink, as well as bigger heatsinks on the VRAM. I just don’t have pictures of that 🙂
Thanks for reading!