Graphics Card Necromancy - Sapphire Radeon R9 390 Nitro

RazorWind

2[H]4U
Joined
Feb 11, 2001
Messages
3,660
It's been a while since we've collectively looked at a dead graphics card, but as I'm sure many of you can relate to, the last couple of months have been a real turd sandwich for me, and I just haven't felt like doing this, but that ends today.

I made a mistake uploading the video to Youtube, so I'll have to post that later, but it's coming, for those who prefer the sonorous melody of my voice... ;)

On the bench, we have this handsome Sapphire Radeon R9 390 Nitro, which came to me in the same lot as the 980 Ti we looked at last time. I was kinda stoked to get this card in that lot, as I can't remember ever seeing a 300 series of any kind in person. I gather that almost everyone who might have wanted one already had a 200 series, and was unmoved by the minor improvements these brought to the table.
CardInitial1.jpg
The seller described the card as totally dead, no fan spin or anything, which is a pretty clear indication of a power issue. Power issues are our favorite kind, as they're basically the only kind it's really practical to repair, particularly without a BGA machine, and furthermore, they're the sort of issue you're most likely to cause through overclocking, which is way more [H]ard than if your graphics card just lost it's will to live.

Our first step is to test resistance on the 12V input rails. We do this first so that if there's a short to ground on any of them, which is a common mode of failure, we know not to power the card up until we've cleared the short.

The near 8 pin connector. I mention in the video that this is probably OK, but hindsight being 20/20, I should have recognized this as faulty. Still, higher is better than lower here, so this is actually not the end of the world.
Far8Pin1.jpg

The far 8 pin connector. Looks healthy.
Near8Pin1.jpg

The PCI-E input pin. Also looks healthy.
PCIEPower1.jpg

Next, we'll plug the card in and apply power, so we can see which DC rails are actually running. We can't probe all of them with the backplate on there, but doing this with the heatsink installed is the best way to avoid the risk of smoking the GPU die. As described by the previous owner, we get nothing on the screen.
NoOutput.jpg

5V rail looks healthy.
5Vok.jpg

3.3V also looks healthy
33VOk.jpg

VCore, though, not so much. This should be in the 700 to 1200mV range. 230 is way low.
VCoreNotOk.jpg

That's everything we can really test with the backplate in the way, so it's time to remove it.
heatsinkrem1.jpg heatsinkrem2.jpg

With the card torn down, we have much better access to do our normal resistance tests. Here's VCore. Looks healthy. Normal should be something like 1.5-4 ohms.
VCoreResistance.jpg

The memory rail looks healthy. 90ish to 150 is what we expect here.
MemoryResistance.jpg

Memory controller also looks healthy. On a 290 or 290X, you'd expect a bit lower than this, but it's in the realm of normal.
MemControllerResistance.jpg

The card's obviously dead, though, so something is clearly still wrong. Let's look at that near 12V input again. A nice thing about this card that I wish more graphics card makers would do is that they've included fuses on each of the input power rails. This means that we're much less likely to burn a hole in the board itself like the 980 Ti did if we have a short somewhere. In the case of that 980 Ti, I suspect the initial failure was actually a power stage, which shorted the input power to ground, and because there was nothing to turn it off, it melted the traces feeding that failed power stage together, destroying the board. A fuse on the 12V input power probably would have saved it.

We know from past experience that the 12V input on these GCN cards should typically be 5K-10K, so that near power connector that read in the millions is suspect. Let's check those fuses.

We've got continuity between the connector and the fuse...
connectorcontinuity.jpg

But no continuity across the fuse. So, we know that this fuse is blown.
blownfuse.jpg

The other fuses appear to be OK. (Far 12V and PCI-E input, respectively)
Far12VFuse.jpg PCIEFuse.jpg

It turns out that the near 8-pin is connected to the top 3 VCore phases, plus the memory power.
Top3Phases.jpg

The far 8 pin is connected to the lower 3 phases.
Bottom3Phases.jpg

If we check resistance to ground on the 12V input for those first three phases, we've got a dead short to ground, so now we need to figure out where it is...
DeadShort.jpg

Anyone care to guess as to how we find the short?
 

RazorWind

2[H]4U
Joined
Feb 11, 2001
Messages
3,660
Alright, video here:

This card's VCore VRM is made up of our old friend, the IR3567B and a series of six IR3553 DrMOS power stages. A DrMOS power stage is a fancy integrated circuit that contains the driver logic and both MOSFETs in a single package. They're attractive to board designers for various reasons, but one of the most important is that they're way more compact. Here's a 290X for comparison (top). See how much less space is taken up by each phase on the 390 (bottom) versus to the 290X? It's enough that they were able to fit a whole extra phase on there.

versus290X.jpg

I always did like the idea of making the MOSFET's casing metal, and using it as a heatsink, though, which is what's going on with the metal-cased MOSFETs on the 290X. That's a pretty cool feature.

Anyway, The IR3553 does not have a particularly good reputation, as it maxes out at about 40 amps of current output, which is not really all that much on a card this thirsty. Youtube is awash with folks who found that an IR3553 failed on their card, and in many cases they were able to repair the damage by replacing one or more of their 3553s. I was all set to remove the top 3 IR3553s, when it occurred to me to give the board one last look over. I started looking at all of the components that could possibly cause a short to ground, when something about this 0805 capacitor you see marked here just... didn't sit right with me. I tried and tried to get a clear photo of it, but this was the best one I got. I should probably invest in a macro lens if I'm going to keep doing this...

badcap.jpg

The visual difference was very subtle, and I don't think it would have been clear without a microscope, but as I dragged a fingernail across it, it would catch on something, as if the capacitor was cracked. In contrast, all of the other caps in that row are smooth. A fingernail glides right across them.

So, I decided to try removing that capacitor first.
Flux on...
fluxon.jpg

I tried my soldering iron first, but the tip I have on it is too small to heat both pads at the same time.
solderingiron.jpg

The hot air wand made easy work of it, though.
hotair.jpg

And it's off!
removed.jpg

With the suspect capacitor removed, I was able to test the circuit's resistance again, and... Yes! A few thousands is what we want to see here. You'll remember we saw 10K on the far 8-pin connector. I'm probing the pads for the removed cap, but these pads are connected to the near 8 pin's fuses and ground, respectively.
shortgone.jpg

Just to be sure, I tested the removed capacitor. Sure enough, it's in short.
shortedcap.jpg

I also removed its buddy, so I could measure it for the replacement. It's 10uF, for those wondering. I know that the other one is also 10uF because the datasheet for the IR3553 specs two 10uF caps in this location.
emptypads.jpg goodcap.jpg

I then ordered fresh fuses and caps from Digi-Key, and called it a day. We'll have to reconvene when they arrive and I can install them. I should note that this short is not necessarily the only issue with this card, but it's the only one that's currently obviousl
 

KATEKATEKATE

Limp Gawd
Joined
Jan 20, 2019
Messages
350
great writeup as always! I enjoy reading about your fixit adventures and appreciate the skill and tenacity that goes into them. wishing you success on this latest repair!
 

RazorWind

2[H]4U
Joined
Feb 11, 2001
Messages
3,660
Video version here:

Moving right along, at length, our replacement parts arrived, consisting of cut tape reels of 10uF 0805 package MLC capacitors and 10A fuses.
newparts.jpg

Here's the dead cap removed, with a quarter, for scale. They're very small, obviously.
0805Cap.jpg

Our first step today is to replace the fuses - the gold colored ones shown here. It's not clear to me why they use different fuses for each input, other than that the code on the other input's fuses indicates they're a lower amperage. So, the current draw from each 8 pin connector isn't perfectly balanced.
fusearea.jpg

I used hot air to remove the old fuses. When they're attached at both ends, you pretty much have to, unless you have a set of hot tweezers.
fusesoff.jpg

I decided to try using my soldering iron first, as it's easier to avoid damaging something else than when using hot air. I'm almost ashamed to post a photo of this, but they're on there, and yes, they're crooked, and also yes, I tagged that connector with the iron. In my defense, the photos make this area look much bigger than it is.
newfuses.jpg newfuses2.jpg

With those on there, it's time to test continuity to the Vin pin on our DrMOSes. This is what we want to see - you may remember that previously, we had about 50M here.
fusecontinuity.jpg

I also checked to make sure I didn't cook the other two fuses, as the one nearest the replaced fuses looks a bit discolored, and it's not flux residue. They're still good, apparently.
otherfusesgood.jpg

For the SMD caps on the back of the card, hot air is the only way to go, especially in this case where there's so much room to work.
Pads cleaned and tinned...
padscleaned.jpg

Flux on...
fluxoncaps.jpg

New caps on...
newcapson.jpg

And done!
newcapsdone.jpg

The solder on these looks better. Note the shinier leaded solder, versus the lead-free solder used for the original parts.
newcaps.jpg newcaps2.jpg

Once again, we need to check for proper resistance. Here, we're testing to make sure that we haven't created a short by bridging the tiny solder pads for those capacitors with a glob of solder. This looks good.
noshort.jpg

It's now time to plug the card in and test. Cross your fingers, and...

Yes! We have a picture!
wehavepicture.jpg

Now that we know the card at least kinda works, let's put it back together. First step was to clean off the flux residue. This took forever; I really need an ultrasonic.
fluxclean.jpg


But I had to take a break to feed Texel...
Texel.jpg

And then I cut new thermal pads. I couldn't find my ruler, so I used this dead 980 Ti. :D As with most cards, it seems, the best replacement thermal pad stuff is the 0.5mm size. The stock pads are a hair thicker, but they're also much softer.
cantfindruler.jpg newpads.jpg

Finally, with the card assembled, I booted into Windows and installed the drivers. As you can see here, it's working fine. I did some benchmarking, and as far as I can tell, the card performs exactly as it should. I get a hair under 9000 in FireStrike, which seems about right for this card and an FX-8350.
benchmarking.jpg
gpuz.gif firestrike.png

:cool:
 

Attachments

Top