When you say the issue appears after time. Did you try additional ventilation/cooling to see if that prolongs the time until the issue appears? Or perhaps it does prevent it completely?

What is the time-frame we are talking about?
Does this issue also appear outside of the operating system, e.g., just hanging around in the bios?

1 Replace the capacitors. Any cheap non-branded polymer ones will do.
1.1 Inspect the board with the lens, ring the resistor assemblies and resistors to the memory.
2. It must be admitted that all manufacturers of that time sometimes did stupid things.
2.1. Measure the GPU core voltage, compare with the reference.
Including the NFB dividers at the PVM controller.
1.6 V instead of 1.4 - what's the problem, it's for stability
2.2 Find the datasheet for the memory and look at the power supply.
DDR1 mutants with 3.3V main power and 2.7V buffer power - no problem, we install them.
We don't want them to go to waste in the warehouse.
2.3 Compare with the images of this particular video card in the reviews.
All the reviews have Hunix memory, do you have Samsung?
Well... it happens.
https://www.techpowerup.com/download/rabit-bios-editor/
In the RaBIt, open all similar BIOSes and find one that is compatible with different timings.
Even if it is "BIOS for Samsung" and you have "Elpida"
Most often, the Radeon 9xxx/X800 series have the same timings in all BIOSes for all memory types. Except for one (often GeCube), which tried and chose something else.
Sometimes they are the ones that fit
3. Accept that not everything can be fixed. 🙁

tehsiggi wrote on 2025-06-30, 09:59:

When you say the issue appears after time. Did you try additional ventilation/cooling to see if that prolongs the time until the issue appears? Or perhaps it does prevent it completely?

What is the time-frame we are talking about?
Does this issue also appear outside of the operating system, e.g., just hanging around in the bios?

No, the cooling of this Asus R9550 videocard (as you can see on the pictures) is passive - not so big aluminium heatsink on top of the GPU die and no option to attach 5 or 12 VDC fan. I replaced the original thermal paste with fresh AS7, but this didn't help.

Time frame is until most of the PCB became worn. Depends of if I use some tests, just DOS prompt or boot to windows. After the card turn off the monitor, the system can not boot 'til the card "cools" again. So, yes it also hangs in the BIOS when the serious problem occurs.

I poked the memory chips and pressed slightly the GPU heatsink when videocard is still "cold" and running and this causes no glitching or any problems. After it gets all worm, without any poking the glitching starts to appear.

That's why I'm thinking - are these problems possible due to failing electrolytic caps?

It could very well be. I'd suggest to measure the key voltages in good state (cold) and bad state (warm) and see if they differ, that could already be a good indicator.

shevalier wrote on 2025-06-30, 10:14:

1 Replace the capacitors. Any cheap non-branded polymer ones will do.
1.1 Inspect the board with the lens, ring the resistor assemblies and resistors to the memory.

Magnifying glass inspection didn't reveal any visible damages or anomalies with this card. I'll have to fully check SMD resistors (and capaciotors) with multimeter - at least for dead shorts or "infinity" values on resistors.

shevalier wrote on 2025-06-30, 10:14:

2. It must be admitted that all manufacturers of that time sometimes did stupid things. 2.1. Measure the GPU core voltage, compa […]
Show full quote

2. It must be admitted that all manufacturers of that time sometimes did stupid things.
2.1. Measure the GPU core voltage, compare with the reference.
Including the NFB dividers at the PVM controller.
1.6 V instead of 1.4 - what's the problem, it's for stability

Fully agree with this. Unfortunately I can't find any circuit diagrams for these R9550 videocards. So I don't know where to measure what.

shevalier wrote on 2025-06-30, 10:14:

2.2 Find the datasheet for the memory and look at the power supply.
DDR1 mutants with 3.3V main power and 2.7V buffer power - no problem, we install them.
We don't want them to go to waste in the warehouse.

I found the datasheet for the VRAM chips - they are GDDR Infineon HYB25D256163CE-40, so good up to 250 MHz and in the BIOS memory frequency is set to 200 MHz.

shevalier wrote on 2025-06-30, 10:14:

2.3 Compare with the images of this particular video card in the reviews. All the reviews have Hunix memory, do you have Samsung […]
Show full quote

2.3 Compare with the images of this particular video card in the reviews.
All the reviews have Hunix memory, do you have Samsung?
Well... it happens.
https://www.techpowerup.com/download/rabit-bios-editor/
In the RaBIt, open all similar BIOSes and find one that is compatible with different timings.
Even if it is "BIOS for Samsung" and you have "Elpida"
Most often, the Radeon 9xxx/X800 series have the same timings in all BIOSes for all memory types. Except for one (often GeCube), which tried and chose something else.
Sometimes they are the ones that fit

Ok, I'll try other R9550 BIOSes, but this card still is with its own stock BIOS. I think these Infineon HYB25D256163CE-40 memory chips are better than Elpida or Samsung chips used on similar R9550 cards.

shevalier wrote on 2025-06-30, 10:14:

3. Accept that not everything can be fixed. 🙁

Yeah, I know. But sometimes I like to try to fix the unfixable 😉

Thanks for the suggestions.

tehsiggi wrote on 2025-06-30, 11:10:

It could very well be. I'd suggest to measure the key voltages in good state (cold) and bad state (warm) and see if they differ, that could already be a good indicator.

This is a good point, thanks! I don't have any circuit diagrams for R9550 videocards, but I'll try to check the voltages on all the VRMs in "cold" and "hot" state. It seems like they are at least 3:

The card is using a switching regulator for the core voltage (VDDC) and linear regulation for MVDDC and MVDDQ.

Check the voltage values at these caps:

tehsiggi wrote on 2025-06-30, 11:27:

The card is using a switching regulator for the core voltage (VDDC) and linear regulation for MVDDC and MVDDQ.

Check the voltage values at these caps:

Thank you very much for this tip! I'll check them when I'm back home and will report later 😉

P.S. I have a proposal, that the problems begin when these two caps became "hot enough" and since you're pointing me there too, the first thing to check will be their voltages in "cold" and "hot" state:

analog_programmer wrote on 2025-06-30, 11:30:

Thank you very much for this tip! I'll check them when I'm back home and will report later ;) […]
Show full quote

tehsiggi wrote on 2025-06-30, 11:27:

The card is using a switching regulator for the core voltage (VDDC) and linear regulation for MVDDC and MVDDQ.

Check the voltage values at these caps:

Thank you very much for this tip! I'll check them when I'm back home and will report later 😉

P.S. I have a proposal, that the problems begin when these two caps became "hot enough" and since you're pointing me there too, the first thing to check will be their voltages in "cold" and "hot" state:

The attachment suspected_caps.jpg is no longer available

For RV360 (RAdeon 9600) Vcore should be around 1.3V
But definitely more than 1.25V and less than 1.35 (this is the ХT voltage)

Re: For all you Radeon ATi RV350/RV280 R9200 / 9200SE/ 9250 /9550 /R9550 / R9600 /9600 /9600XT lovers DX9 , schematic h

shevalier wrote on 2025-06-30, 12:06:

For RV360 (RAdeon 9600) Vcore should be around 1.3V
But definitely more than 1.25V and less than 1.35 (this is the ХT voltage)

This must be the voltage on the positive legs of the two "suspected" capacitors, right?

shevalier wrote on 2025-06-30, 12:13:

Re: For all you Radeon ATi RV350/RV280 R9200 / 9200SE/ 9250 /9550 /R9550 / R9600 /9600 /9600XT lovers DX9 , schematic h

Great! I didn't knew there's a thread with the diagrams for these Radeon 9xx0 videocards. I'll check them out.

Thanks for all the info!

analog_programmer wrote on 2025-06-30, 12:18:

This must be the voltage on the positive legs of the two "suspected" capacitors, right?

Thanks for all the info!

Yes.
If you don't like fiddling around with two probes at the same time, you can use the slot bracket as your ground point (negative terminal) and then probe only using the positive probe from your multimeter.

tehsiggi wrote on 2025-06-30, 12:22:

Yes.
If you don't like fiddling around with two probes at the same time, you can use the slot bracket as your ground point (negative terminal) and then probe only using the positive probe from your multimeter.

Yep! I always measure voltages this way - negative probe stuck to some suitable "common ground" point while using only positive probe to poke at the PCB elements 😉

I'll check all these regulated voltages when I get home later. Thanks for the all the tips and the information.

analog_programmer wrote on 2025-06-30, 12:18:

shevalier wrote on 2025-06-30, 12:06:

For RV360 (RAdeon 9600) Vcore should be around 1.3V
But definitely more than 1.25V and less than 1.35 (this is the ХT voltage)

This must be the voltage on the positive legs of the two "suspected" capacitors, right?

Honestly, I wouldn't measure anything.
I'd just take a couple of polymers from a donor AM3/LGA775 (or a video card), something like 470-820uF* 3.3-16V (in any combination, as long as they match in diameter)
Not sporty?
But practical.

analog_programmer wrote on 2025-06-30, 12:28:

tehsiggi wrote on 2025-06-30, 12:22:

Yes.
If you don't like fiddling around with two probes at the same time, you can use the slot bracket as your ground point (negative terminal) and then probe only using the positive probe from your multimeter.

Yep! I always measure voltages this way - negative probe stuck to some suitable "common ground" point while using only positive probe to poke at the PCB elements 😉

I'll check all these regulated voltages when I get home later. Thanks for the all the tips and the information.

For low voltages and high currents - this is incorrect. The error can reach 10+ percent.
Measured with both probes on the terminals... of ceramic capacitors as close as possible to the GPU itself.
But for a quick check - this is inconvenient.

shevalier wrote on 2025-06-30, 12:31:

Honestly, I wouldn't measure anything. I'd just take a couple of polymers from a donor AM3/LGA775 (or a video card), something l […]
Show full quote

Honestly, I wouldn't measure anything.
I'd just take a couple of polymers from a donor AM3/LGA775 (or a video card), something like 470-820uF* 3.3-16V (in any combination, as long as they match in diameter)
Not sporty?
But practical.

I'm an electrician (not so much into electronics, but still trying) and I always prefer to do some basic measuring when possible. Just to be sure, like in this case 😉

And what if the GPU voltage is out of specs (less than 1.25 V or more than 1.35 V) - do you think that simple recapping will "cure" the problems?

And why polymer caps instead of new low ESR electrolytic caps?

Since we're talking bulk caps here, at the output of the switching regulator on a card that draws at max 25W, this measurement should be sufficient. If the caps are bad, the voltage issues should be visible there already. For a RV350/RV360 based part the expected current taken by the GPU is around 7A.

And why not looking if a proper measurement can reveal something? I think there's more joy in analytical approaches than just exchanging parts and be good with it.

If it is way lower / way higher (which I don't expect) it'd be due to the caps missing capacitance and thus the switching regulator not being able to keep up the voltage due to the high current demand. Also the ripple on the output of the regulator will be severely increased.

Guys, don't argue on this. I'll measure the voltages for sure. I've learned my lessons during years to not mess with the things 'til I get enough information, so I can isolate the problem in the smallest margins possible.

Unfortunately I don't have oscilloscope at home to see the ripple on the output voltages of these VRMs. But I fully agree, that a simple voltage measuring will be enough to be sure if the capacitors are failing, without the need to desolder and put them into one of those cheapo chineese component testers.

analog_programmer wrote on 2025-06-30, 12:37:

And why polymer caps instead of new low ESR electrolytic caps?

So that you don't have to resolder them again in a few years. Polymer capacitors degrade very slowly, and with a power of 20 years ago (20-50 W) they will last forever.

And what if the GPU voltage is out of specs (less than 1.25 V or more than 1.35 V) - do you think that simple recapping will "cure" the problems?

In order for the output voltage of a switching power supply to drop, only the name of the capacitor must remain.
The main problem is voltage ripple.
Because when the voltage drops, it becomes below the minimum for switching logic elements at a given frequency.
Ripple Voltage = Capacitor ESR * Current_Change
For example, if we have a capacitor with an ESR of 0.45 Ohm, when the current changes by 1 Ampere, we will get a ripple of 0.45 V.
And with a constant change in current of 1 ampere, (1*1)*0.45 ~ 0.45 W of heat will be released on the capacitor. It will be stupidly hot.
Therefore, a tester alone is not enough, nor is a simple capacitance meter.
You either need an ESR meter or just replace the capacitor (just in case).

tehsiggi wrote on 2025-06-30, 12:40:

And why not looking if a proper measurement can reveal something?

That's why I settled on 0.44 ohms. 😀
https://youtu.be/-C8kqmCd-m4?si=XNjus2GuddaL2rGd&t=714

So far, so good. Or good to be not good 😀 Let's see what the multimeter shows:

First I measured the cap's terminals resistance to ground - just to be sure they're not shorted:
1 - gnd = 5.7 Ohms (suspicious low resistance, but still i don't know what's the switching voltage regulator circuit)
2 - gnd = 0 Ohms
3 - gnd = out of multimeter range (too big resistance)
4 - gnd = 0 Ohms
5 - gnd = 89.5 Ohms
6 - gnd = 0 Ohms

1 - 2 = 5.7 Ohms (as expected)

And here's the revealing voltage part:

I fired-up the system and immediately measured the voltage values at the points when the card was still "cold":
1 - gnd = 1.205 V (ooops - too low, eh?)
3 - gnd = 2.675 V
5 - gnd = 2.578 V

Then I waited the card to "heat-up" and start glitching and the measured values were same.

Then finally, after the monitor switched-off and the system hanged, I took another measured values with "hot" card:
1 - gnd = 1.204 V
3 - gnd = 2.674 V
5 - gnd = 2.578 V

So no difference, but obviously GPU voltage is not enough.

After the card was c0oled again, I repeated the experiment, but this time I measured the voltages on both terminals of the capacitors (as shevalier insisted) - 1-2, 3-4 and 5-6. And there was absolutely no difference.

So. I think it's time for recap. At least for the two "suspicious" capacitors on the GPU core voltage circuit. I don't know if the other two voltages from the linear VRMs are good enough.

What do you think, guys?