Are you failing prime with small FFT's or are you failing in Blend?
Does Intel Burn test / Linpack also fail ?

Try removing the northbridge heatsink and the push pins mounting it and the mosfet heatpipes (you will need to remove the motherboard to carefully use tweezers (do NOT damage the pins by pushing too hard) to gently push them in), then use goo-gone or alcohol to remove the caked on "adhesive" stuff on the northbridge and on the back of the heatsink, and be very careful when you do this. You can take a knife if you want, and CAREFULLY, with the help of some alcohol, very slowly help loosen the caked on adhesive compound on the northbridge; this is safe but you have to be very careful; your job is to get the compound removed, not scratch the northbridge. When you're finally done, apply some arctic Ceramique (I would not recommand AS5 because there are too many exposed resistors around) or even some generic white silicone compound, and then remount the heatsink.
If you have a small fan (they're less than $10) you can screw that into the northbridge.

Then reboot and set the mch voltage to 1.65v then if that's not enough, try 1.75v. Increase FSB termination voltage to 1.30v, and then try 1.40v only if you need to. Do not go to 1.50v.

You can also see your board revision when you're doing this; its written above the RAM slots. All 1.03 and 1.04 boards can go to 1.85 vmch, but it's a bit risky to go that high on the mch, even with it actively cooled; 1.75v should be ok with active cooling.

After you do all of this, Do some memtest86+ loops, and especially focus on test #5. See if you can get that to pass 10 loops; if you can, then your memory and northbridge should be stable.

Then run prime95 small FFT's and try a couple of 5 loop passes of Linpack if you want.

Something to keep in mind: Raising the FSB termination voltage also increases the vmch voltage slightly; about .06v for ever +.1v of FSB. Tested by someone using a DMM. I do not know if this applies to boards with 1.65v max vmch.