radeon vram swap...

[kiaas]

I'm currently working on doing a VRAM swap on some videocards, planning on putting 8Gbit chips on an Rx 470, donated from a dead 480 with a chipped die, and then the 4gbit chips from the 470 onto a 7850. I've been trying to figure out what in the vbios needs changed/copied over, and have most of what is documented in atombios.h figured out namely the vram_info that's 33bytes not including part-number, and memory timings for vram-module-info v7 and 45bytes not including part number in v8. I'm not sure what else has to be copied over/converted, and even in those there's a few unclear values. I did find some fairly in-depth register guide for some terascale cards, specifically a pdf on xorg's site, but I doubt that'd carry over to gcn, and the stuff for GCN cards seems to leave out a lot of those details.

Anyone got any guidance for what other details will be needed, and how to interpret some of the board-specific things?

[kiaas]

I doubt you can solder surface mount components. You need a desoldering tool too.

I have already done it and reballed a number of chips from dead cards to get experience in the work. I've been working on figuring out getting reballing right consistently, I think my stencil or balls aren't made to adequate tolerances and that has been causing me problems on some practice chips, I have different ones on the way, to make sure of it.
edit: I am using the wrong hot-air source, heatguns work but lack of temperature control and high airflow will more readily degrade things or blow off small SMDs, the latter can be helped with kapton(polymide) tape and foil shields, the former is more difficult. 60-80 dollar hot-air stations on amazon are good enough to fairly reliably do smaller mass components (like vram modules.), though those stations sometimes have the wiring connected wrong internally so should be checked and fixed if it is to prevent burning anything, but at least the actual components tend to be fine. BGA rework is new to me which is why I'm doing a boat-load of practice first, but I've done tsop with my heatgun successfully, and qfp and qfn packages with real hot-air reflow stations.

[kiaas]

There is no sense to put 8GB vram to RX 470. 4GB is enough for 4K gaming. I have measured VRAM usage with the Rise of the tomb Raider game at 4 K and RX570. The game uses max 2GB of VRAM. You can break your working RX470 if you make a mistake and probably you will do unless you have done a working GPU card modification before.

I have use-cases where 8GB of Samsung memory is much much preferable to what's on it.
I already planned for testing it on unimportant cards first, and I've got some exceptionally low-value cards for it, a 10 dollar hd 6670 for one, and ones demonstrating problems (green sparkles, either pre-existing damage to vram or need for reflow already, if those come out equal or better, it'll be a success for my purposes.)

The only part of this that is new to me is the reballing, pulling BGAs isn't much more difficult than QFP or QFN, reflowing BGA is probably easier than QFP and QFN, both of which I've worked on as actual work. Can't really buy GDDR5 new with balls already anymore, mouser and digikey haven't had stock available in small quantities for a while, so I have to learn this extra step. Some chinese sellers do have stock of chips but I'm not sure if it's used and reballed or extra stock they got their hands on, and it isn't that cheap so this seems like a better option.

My progress has been: Taking chips off of a broken hd 5670. There were 8 of them, in clamshell-mode on the pcb. All came off quick and easy without damage. Cleaned them with flux and desoldering wick with my iron, worked fine. Attempted to reball using normal procedures I've read and seen, but had a regular failure of balls not getting attached to pads, and some getting stuck in the stencil.. Only managed to get it to work on half the chips with multiple retries.
Thinking it might be something to do with the stencil or balls, I changed technique to putting kapton on the top of the stencil, filling the holes from the bottom with balls and putting a thin layer of flux to hold them in, fluxing chip, applying stencil and heating. Worked much better but still have problems that seem related to the stencil, as there seem to be 2 balls that keep getting stuck. this is where I've paused practice on undamaged chips to not damage any more, as I wait for new stencils. I am continuing on chips that had pads lifted already, however, to get all but the padless spots to stick without a second hit of heat.
I may make something to compensate for stencil/ball problems where they get stuck even with other stencils coming in that hopefully won't have these problems, and I will be buying a better hot-air tool, as I see the price on some that have pretty good reviews have come down a lot, I have no reason to not get one now.
edit: oh, also the vbios editing. That's new-ish, I've edited things in vbioses before with hex and some tools for doing other things with them, but that's what I was here asking about for things beyond just timing or clock changes as there's only so much documentation I've been able to find, mostly atombios.h in the linux kernel source.

[kiaas]

I expected it to work just like reflowing anything else, and the fact that when only 1-2 balls don't stick I can manually place them on the pads and reflow with a 100% success rate on that second hit of heat is more evidence something was made to the wrong tolerances, I can deal with that by getting more samples that hopefully will be closer to spec, among other work-arounds. if I'm missing something in the vbios, well that'll be a little harder to see what's wrong and figure out how to fix it.

[kiaas]

Getting closer to needing to confirm what parts from atombios will need modified for the end-goal.
Got 3 new stencils in. 2 of them had the same problem, to a lesser degree than my first... but I got 10 chips reballed successfully in a row with the 3rd. Next step, reball vram on some extremely low value cards just to make sure I'm staying within reflow tolerances of the chips, if that fails repeat but with a better hot air station. THEN, I can move on to actually changing memory with vbios mods to support it on something.

[kiaas]

first attempt on the 10 dollar card to pull, reball, reflow it back onto the card went as I expected. I didn't "kill" the card, but did damage it, which is why I started on a 10 dollar card. Hot air station will be in tomorrow for better temperature controlled work and I'm pretty comfortable with the actual reballing process now. I'm not sure if I'll have it fully figured out on the next try to actually have a card survive 100% functionally, but I should be able to succeed by the one after that.
Getting replacement VRAM reballed by chinese scrappers to put on the 10 dollar hd 6670, and testing that before trying to pull and reball it again myself will be one of these steps before moving on to things I'd like to work for more than proving my success.