RX 470 Mining back from the dead - zombie edition

[katapaltes]

So here's the TL;DR question, but I'll put the details below: If I removed one of four inductors (chokes) from my GPU's VRM, should the core now be running at only 300Mhz instead of 1206Mhz? I was hoping I could run using three of the four power phases if I seriously underclocked/undervolted the GPU.

Details: I bought a Sapphire RX 470 8GB mining edition (DVI-only) on eBay months ago. It had been used for mining and had a mining BIOS on it. The white sticker on the fan shroud says it's a "11256-57." The card turned out to be DOA, with both fuses for the 8-pin PCIE power plug dead. I contacted the seller and he gave me a refund and allowed me to keep the card. Thus began my adventure.

I know little about electronics, but wanted desperately to learn. At this point I've bought and used for this project: A cheap hot air rework station, .5mm Chipquik solder, flux, solder wick, Littelfuses from China, and Fast Chip removal alloy. I've spent more on the card than it's worth, but it's all about life-long learning so it's worth it. Oh, and I bought another used RX 470 4GB mining card (model 11256-58) from eBay so that I could compare voltage, impedance, etc. between the two cards. I replaced its mining BIOS with 193398.rom from this site and it works great.

After watching hours and hours of videos (some of them repeatedly; some of them in Russian *) and reading forums like this one, I guessed that at least one DrMOS was dead. It was then that my nightmare began. I could not for the life of me remove the damn FDMF3035 DrMOS ("integrated MOSFET plus driver power stage") with my $50 hot air station and limited skills. In fact, I accidentally dislodged a couple tiny passive components with the hot air while the DrMOS would never budge. I even tried using a heated frying pan and removed the card's bracket to help with the heat sink effect.

I read that I could disable one phase entirely and that the card should run without it. I saw that removing the through-hole capacitors wouldn't help in troubleshooting and that I'd have to remove the SMT choke. Again, hot air didn't do the trick (for me) so I bought the $20 tube of Fast Chip and HOLY CRAP THAT STUFF WORKS. In the photo below, you can see the dislodged choke, two disturbed passive components both hanging on for dear life, and the real culprit, a scratched-up FDMF3035 DrMOS.

With the choke removed, all my impedance testing was perfect. I installed the card in my throw-down PC and got fan spin, and then Windows launched. I then flashed with the closest BIOS I could find from this site (209467.rom) as the installed mining BIOS was whack. Aside from my TL;DR question above about running on three out of four phases, is there any hack where I don't have to replace the DrMOS? Is the only path forward to replace that stubborn IC?

* I turned YouTube's auto-translate on and it translated one of the Russian words in the GPU repair video to "Tanya." I sent the video to my friend Tatyana (a.k.a. "Tanya") and asked her if the translation was correct.

 

[katapaltes]

Update: At this point I've edited the video card BIOS to make it run higher than its 300Mhz limp mode. I don't think I'll ever replace the failed DrMOS, so I went ahead and completely removed the tiny passive components that I accidentally dislodged just to be safe.

Anyway, setting Power State 0 (zero) on the card to 751Mhz (or whatever feels "safe") makes the card run at that speed on boot. I would love to know a way to do it using software in Windows, but with a power phase removed, that doesn't seem possible.

The pic below is of the GPU clock stuck at 300Mhz while running GPU-Z's Render Test where it normally would run much higher. After editing the BIOS to run Power State 0 at 751Mhz for example, it runs at that speed (and never below that speed, because it is Power State 0 after all...). I don't think the Wattages highlighted below are accurate. I think some calculation must be thrown off by the missing phase.

 

[msimax]

nice

 

[Caring1]

The pic below is of the GPU clock stuck at 300Mhz while running GPU-Z's Render Test where it normally would run much higher. After editing the BIOS to run Power State 0 at 751Mhz for example, it runs at that speed (and never below that speed, because it is Power State 0 after all...). I don't think the Wattages highlighted below are accurate. I think some calculation must be thrown off by the missing phase.

Quick question, where is the GPU-z render test found?
I've pawed over the latest edition and can't seem to find it.
I've just picked up a cheap RX570 that seems to be stuck at 300MHz so want to test things.

 

[katapaltes]

Thought I would include a few more details and a pic of GPU-Z with the BIOS modded with Power State 0 (p0) set to 751Mhz. Nothing magical about that speed afaik; It's just a speed I copied over from Power State 2 (p2) along with p2's milli-Volt value. Speeds on this card range from p0 to p7 (300Mhz to 1206Mhz).

So now the card boots up at 751Mhz and stays there. Speed doesn't ever go UP (whether p0 is 300Mhz or p0 is 751Mhz, card won't allow itself to go above p0, I'm assuming because it knows something is wrong) and it does't ever go DOWN (I'm assuming because p0 is always the lowest speed by definition). A normal card like this one starts out at 300Mhz and goes up and down.

So I guess I have to pick a number and set the card to run at that speed. I flashed the BIOS to set p0 to 1019Mhz and it ran fine and played games. But later it didn't start back up right away and I had to fiddle with it. I may have to pick something more modest that's closer to the 75% power capacity I'm working with.

I imagine the card is using more power at this fixed, higher GPU clock. But the temps aren't bad and fans slow, then stop at idle. I had to run atikmdag-patcher against the video card driver or it wouldn't recognize the modded BIOS. I also saw advice that Polaris BIOS Editors were often infected with malware, but the copies of PolarisBiosEditor-1.7.2 and 1.7.0 I got from github checked out OK at virustotal.com.

I've read that undervolting is a good way to make life a little easier for your video card. Anyone have an opinion on doing that with a card that is missing a power phase? Would undervolting be the exact OPPOSITE of what I should do?

Quick question, where is the GPU-z render test found?
I've pawed over the latest edition and can't seem to find it.
I've just picked up a cheap RX570 that seems to be stuck at 300MHz so want to test things.

On the main GPU-Z screen, look on the right-hand side for a question mark ("?"). That's it.

 

[katapaltes]

Update just for fun: Looks like the "limp mode" has also lowered the card to PCIE 1.1 @ x16 even under load. This is not a huge concern for me, as an RX 470 used for gaming shouldn't need a ton of bandwidth. The only problem is if I put it in Crossfire with my other 470 I bought for this project, and then I'm down to 1.1 @ x8 - and then add Crossfire traffic going over PCIE to the equation. I did try Crossfire and it seemed to work reasonably well; was getting maybe 130FPS in Prey 2017 on the highest settings.

Does anyone here know a way to set the PCIE version in the card's BIOS? I have the option to force version 1.0/2.0/3.0 in the PC's BIOS, but it didn't help in this case. GPU-Z still reported version 1.1.

Whoops... I left the Render Test running accidentally as I was writing this and then I saw in GPU-Z that it had *finally* somehow bumped up to 3.0. I shut down the Render Test and now it won't drop back down to 1.1. Minutes later and it's still hanging out at 3.0 with just a browser open. I'm going to hope that reading is accurate and move on, even though GPU-Z to this day reports one of my PCIE 2.0 slots as 3.0...

Regarding my previous question about undervolting, I am running the card at 926Mhz and 800mV. 926 is the speed of the reference RX 470 (not bad for a card missing one of its four power phases) and 800mV is around a 25mV undervolt at that speed.

 

[katapaltes]

It's been a good weekend. I'm now running the card at full speed (1206MHz), still with one of four power phases disabled.

I sent a detailed request to a local electronics repair shop who has good reviews, and they wanted $120 to resolder the needed components. I decided against since the card wasn't worth half that amount and it was still somewhat usable at 75% performance.

Turns out the PCIE Gen 1.1 limit I mentioned above was caused by the Adrenaline 2020 drivers. GPU-Z showed Gen 3.0 when I downgraded to version 19.6.3 from 2019 - or if I was running Microsoft's Basic Display Driver. (An article at Igor's Lab tipped me off to the 2020 driver behavior. "Whoever sets the PowerLimit, the clock or voltages too high will be punished with a GPU clock of 300 MHz which can no longer be raised.")

At that point, it was worth trying for more so I started raising GPU core clocks and voltages slowly using Polaris BIOS Editor, rebooting, and testing for stability. I was able to move from my cautious setting of 926MHz/800mV up to 1156MHz/940mV, a 230MHz increase, and within spitting distance of the 1206 default top end shown in my BIOS. I was pretty happy at that point, the only problem being I was using 75W at idle instead of 55W as the card was set to run at 1156MHz full-time instead of dropping to 300MHz as needed.

Curiously, the card needed about 50 to 60mV more than stock at any given frequency/Power State. I'm sure there's a good reason for this with the missing power phase, but I can't guess as to why.

Again with Igor's advice in my head and knowing that the 300W-plus readings in the screenshots above were bogus, I used PBE to set some ridiculously high limits for TDP (W), TDC (A), and Max Power Limit (W). HWInfo was showing all three of these at over 300 Watts or Amps apiece, so I somewhat randomly chose "369" and put that into an otherwise-unmodified version of the BIOS.

The result? I'm able to run Adrenaline 2020 drivers with PCIE Gen 3.0 and the frequency goes up to 1206 as needed. I am hoping the card can last long enough for us to get through this GPU shortage, at which point I'll buy a new card. I'm going to try for some mild overclocks next, but I may be at the limit of what my gimped card can supply.
https://www.igorslab.de/en/red-bios...en-more-stable-overclocking-navi-unlimited/3/

 

[CalzRei]

Nice! Like they said, keen eyes and dexterity comes a long way. Dealing with a weird card myself rn, boots to windows just fine but launching a 3d app will immediately crash it

 

[delshay]

This is how I would remove that inductor,. (Extremely easy for me)

Don't use a hot air gun, wrong tool without a preheater, & even then I would not use A hot air gun or a full infrared system which I have.

This is how I would do it, very easy.

Get a high wattage soldering iron with a big tip & melt the solder at one end. Then get a very small flat screwdriver & lift it ever so slightly away from the PCB. If you can get the screwdriver between the inductor & PCB, this is better as you can leave the screwdriver there & remove the soldering iron. When cooled down, remove screwdriver. ...NOTE I'm talking around 2-3 mm gap between inductor & PCB.. Now the other side of the inductor is straight forward to de-solder & remove component.

Note that this only works with SMD inductors & you can use a pair of pliers also to create a 2-3mm gap at one end (de-soldered)).

This is what I do here when I don't want to heat-up the PCB with an infrared system..