I have junk AT PSU which only outputs:

5V cable: 4.84 V
-5V cable: -4.64 V
12V cable: 11.65 V
-12V cable: -11.17 V

however, the Power Good cable stays at GND, indicating that the voltage levels are not in spec, thus not letting the motherboard turn on. What did your other voltages look like? If they were as bad as mine, your MB shouldn't be powering up.

Side note: caps aren't bulging. Anyone know which components need replacing in this AT PSU (caps, regulator, inductor)?

I just recently acquired one of these CPUs and decided to push it to 100MHz also.
I must say I have not read through this entire thread so if someone else has already documented this, I apologise.

Firstly let me clear up some misconceptions regarding the voltage regulator vs the motherboard's voltage regulator.

On Socket 3 there are two sets of VCC pins, one set is VCC5, the other is VCC3. VCC5 is always connected to the +5VDC directly, where VCC3 is provided via a regulator (adjustable or not) from the motherboard.

The POD83 internally bridges these two sets of pins, bypassing the onboard regulator. You can prove this simply by measuring for continuity between +5VDC and the motherboards regulator output. When you insert the CPU, you will find a short immediately develops between them. This explains why early in this thread, it was mentioned that shorting the in-out pins on the regulator resulted in ~4.9V.

Regarding the diode mod here, I really can't understand why this is in use. The onboard regulator is clearly a 3 pin fixed voltage regulator. The only difference between a fixed and variable regulator is the resistors that set the voltage on the ADJ (adjust pin) are internal to the package. You can always adjust the voltage simply by adding an external resistor between the ground pin and ground. Conveniently this pin is the one right on the edge of the CPU package.

So how I modded my CPU:

1. Solder a wire onto the middle pin of the regulator so I can probe the voltage it's running at with a multimeter (this is the white wire).
2. Lifted the regulator pin closest to the edge of the CPU
3. Soldered a multi-turn 1k potentiometer between the lifted pin and the pad on the CPU package

Using this I am able to dial in an exact voltage, which does not drift under load as the regulator is simply performing it's intended function.
Raising my CPU from the factory 3.5v to 3.6v has made it completely stable at 100MHz.

I wasn't aware it was so simple to mod a fix-voltage VRM. The diode mod started as a quick experiment, and since it worked well, we just stuck to it. It delivers around 4.15 V and so far, we haven't run into a POD that couldn't do 100 MHz at 4.15 V.

I guess you'd adjust your trim pot's size based on what internal resistors were used. With 1 K-ohm, what voltage range are you allowed? What software did you run to determine stability at 3.60 V?

I have not done any extensive testing yet, I did however have to bump it up to 3.65V as I was getting the odd invalid opcode issue launching Windows 3.11.
I will work on getting some more software onto the system so I can properly test it and verify.

As for voltage range, 3.50 through to 4.0v seems doable, any more and I think it's hitting the limits of the regulator, or I need a higher value pot (it's hard to tell when you hit the end of a multi turn pot).

I think you might find that you don't need such a high voltage if you use this method because the voltage is stable and doesn't vary with the drop in the diode junction under different loads. More testing is needed to verify though.

Happy to report that it's solid stable at 3.65V, verified by running Quake timedemo in a loop for several hours.
Watching the CPU voltage, it dips to 3.64V while under a heavy load.

Edit: Ok, no it wasn't stable, once I went to a higher resolution of the timedemo it couldn't complete a run.

Watching the voltage over the longer run, it drops and continues to drop until it goes too low and crashes. The problem is the regulator is getting too hot and self-limiting to protect itself, a common behaviour in linear regulators. I found that if I pushed a thermal pad into the gap to allow the regulator to conduct it's heat directly into the heat sink above it, it helped, and I am now able to complete the quake time demo at 640x480.

While I have it working now, I think I might investigate building a off-board regulator that replaces the one on the CPU, I am curious to see how low the voltage can be if the rail is stable.

I find running numerous Windows 9x apps to be more telling in terms of stability than DOS Quake.

At what dropping voltage did the CPU crash? How hot did the regulator get?