A water temperature (or ambient) would be handy. If I remember correctly, lm90 + Tbred has an error of +6/-4, nearly all of which is systematic error (the offset will be constant for a given lm90/Tbred pair). You can roughly determine that error by plotting delta-T (CPU minus water) vs multiplier (keep Vcore, FSB, etc. constant).
Prime95 isn't a good heat loading program, and running two instances just increases OS overhead and risks cache thrashing. Use
K7burn.
A good possibility for the excessive heat increase is that your CPU is at the edge of stability (a distinct possibility at 1.92 with a Tbred A). At the lower voltage the CPU is doing less work than expected due to cache ECC errors and is idling during the subsequent calls to memory to fetch the correct values. At couple more voltages above, between, and below your two data points could tell the tale.