Be careful using the XT diags, the XT cards were not just a refresh they redesigned the PCB used in the earlier 9700 Pro and 9800 Pro cards to resolve issues related to the die shim and to add in thermal sensors, some of it will be similar but the rest may differ a bit.

I do hope you get this old girl up and running though, far too many 9700 Pro and 9800 Pro cards are DOA or have issues.

The die wakes up, presumably you've replaced the mosfets and now something gets a power good signal.

Does the memory get a voltage?
Bios readable/flash able?

If you haven't replaced the mosfets, and instead blindly pumped 9v into the core... I would expect it to be very dead.

MOSFETs have not yet been replaced, footprints are left empty so the rail voltage can't reach any device at the buck output side.
I only supplied from the AGP PSU side (+12V, 5V, 3.3V rails), as if I replug it on the motherboard. So all should be safe.

Oh. Right. So the slot 12v circuit is short, but the molex circuit is normal.

Weird because you'd expect an and gate to trigger that circuit to start. Tough I suppose it could be 5v and 3v and it doesn't care about 12v for some reason.

You did look on both sides for hot components, right? If yes, pull the mosfet driver. They often fry when fets die.

Oh. Right. So the slot 12v circuit is short, but the molex circuit is normal.

Not exactly but I might not have been clear enough!
I think that the rails coming from AGP slot are fine.
Among circuits and functions supplied by these rails, there is this dual phase buck. Simply speaking, one phase is supplied by the AGP slot whereas the other one is supplied by the Molex plug.
I removed the 4 MOSFETs at the heart of this dual phase buck (only one MOSFET is failed) but I still have a short circuit at the output of this power stage, "behind" the MOSFETs, on the smoothing inductors side.

Since this part of the circuit is isolated from the rest (MOSFETs are removed), I can feed the AGP rails (+12V, +5V, +3.3V), there are no shorts. In these conditions, the die seems to come back to life. Temperature rise is not exploding and the drawn current is not great, not terrible.

But I still have to find the short on the buck output side...

Hmm.
Different to these days, but different era, I guess.
These days, the slot power would do ancillary things mostly, maybe memory at most.
Still, it "lights up" as it were, so you'd expect it capable of working.
Check vccmem, and bios next.

But ideally, you find the short.
Again, it's probably the driver IC.

If Techpowerup is right:
https://www.ixbt.com/video2/images/r9800pro/r … o-scan-back.jpg

It's this
https://www.alldatasheet.com/html-pdf/1050686 … /1/ISL6523.html

These days, the slot power would do ancillary things mostly, maybe memory at most.

True.
However I read somewhere that the 9800s can boot without the Molex plug connected but it will display some message signaling lack of power.

But ideally, you find the short.

Sure!
It's incredible... I sourced 5A into this shorted output and absolutely nothing spotted on the PCB with the IR camera.
Yet I could clearly see the power supply's cables heating!!!!

Heh. See edit.

How do you know that it's a "dead short"? I'm not familiar with that 9800 PRO model, but the core itself can exhibit a very low resistance, which might mislead you to think that it's a dead short.

It's this

For single phase bucks (there are 3 of them on my 9800 PRO), it's ISL6522CB.
For the single dual phase, this is the SC1175CSW, on BOTTOM side, just under the 4 MOSFETs.

I have listed all the relevant SOIC chips for TOP side :

And for BOTTOM side :

(red dots are marks for suspicious devices... I still have to check the STS8DNF3LL on bottom)

How do you know that it's a "dead short"? I'm not familiar with that 9800 PRO model, but the core itself can exhibit a very low resistance, which might mislead you to think that it's a dead short.

I'm quite sure because my lab PSU indicates almost 0.5V for 5A sourced, which is 0.1ohm, the resistance of my wiring (and I see it heating with the IR camera!)
There is no hot spot on the board, i.e. no dissipated power in a (significant) resistance --> dead short.

As a reminder, I removed the electrolytic capacitor (3rd one, the lowest, near the 2 "Falco" power inductors) to supply current directly at this point. This is where I have the short.

Are you sure you haven't accidentally tested ground to ground?

I will double check but I remember that I already have the short when I removed the 4 MOSFETs, and probed at the switching node...

Possibly one of the most common problems for the Radeon 9700 and 9800 cards was bad solder joints under the core. That alone could be enough to result in no output while the core itself still gets hot.

Do you have an AGP tester with the LED's that indicate connection from the slot to the core? (Search AGP tester on Aliexpress and you will usually find a combo tester for PCI-E and AGP)

Perhaps you could replace all the parts that you have removed and then do a reflow of the core and see if anything changes?

If you're going to reflow it, might as well reball it.
Op sounds like they might be skilled enough to try

myne wrote on 2025-01-10, 05:35:

If you're going to reflow it, might as well reball it.
Op sounds like they might be skilled enough to try

Agreed but reballing is considerably more difficult and requires significantly more skill. It can be especially more challenging without a stencil.

Personally, I do not have the skill to reball a GPU core at this time, nor do I have stencils for any GPU cores (especially for old cards like the 9800 series) but I can do a reflow and have been successful in repairing a few cards, including a Radeon 9700 with that process.

I use kapton tape and foil (to protect components that I don't want to heat with hot air), a preheater, hot air station and flux. (Not an oven)
Older cards that use leaded solder don't require as much heat, so they're definitely easier to reflow.

I think it's worth a try.

byte_76 wrote on 2025-01-10, 06:58:

It can be especially more challenging without a stencil.

It's really not. After a bit of experience it takes ~ 20 - 30 minutes to place ~ 500 balls. 😀

I'll just leave this here

PS. Whether or not to display a warning about the lack of external power is determined by the BIOS
In any case, the card will not work without Molex.

In my case, M1 to M4 removed, measure 0 ohm at Vcore node.

Is the core the unique load at this Vcore point? Or other auxiliary switching circuits deriving voltages from Vcore?

It’s normal to measure 0 ohm for vcore unless you have a very sensitive multimeter. (or a milliohm meter)