Hey Ivan,
Best bet is to give us a call. But I'll respond with my current line of thinking in case you aren't able to do so.
Plotting, i.e. rendering speed, is the weakest link on most systems. Though with high loads this typically results in plotting lag or the plots being less responsive. The rendering and math engines run in different threads and are loosely coupled, thus have limited affect on each other. If the CPU load is high, then graphics can exhibit some lag in some environments. But graphics lag should have little to no affect on the math engines. Worst case if the load is high enough, an overall performance hit could be encountered, but that shouldn't impact the accuracy of the math involved; but merely slow it down.
A +/-15dB variance is massive. I find it interesting that it's only below 2k. How wide is the variance? We talking ripple or multi octave bump? Is there a pattern to the variance? I ask because I'm thinking crosstalk or some other driver anomaly. I'd like to attempt the same test using a different device, if possible. That should assist in discerning if the device or driver is the cause.
That's the best I can do with the information at hand. I'll post more if anything else comes to mind.
-A