As some of you may have noticed, I recently upgraded an oldish Socket 754 single-core PC with a dual core X2 3800+ Socket 939 setup.  I bought a nForce3 motherboard so that I could use my AGP video card and DDR memory, and all-in it was a pretty cheap upgrade.  Thankfully, the fantastic Arctic Cooling Freezer 64 Pro heatsink/fan unit (which can be purchased for $20 or less online) also works on s939, so I already had all the pieces in place for some good aftermarket cooling and some aggressive overclocking!

1.525vcore seems to be the sweet spot for this CPU, and core temps are staying in the low 60s, so I feel pretty good about that.  I noticed though, as I dropped the HTT multiplier and RAM speed in order to eliminate those barriers to the CPU overclock, my FAH output (as monitored via FAHMON) didn't increase as linearly as I expected.  As the CPU speed ramped up, FAH output lagged behind.  I settled on a maximum CPU speed of 2.7ghz (with a FSB speed of 270mhz).  2.75ghz was "Windows stable", but the SMP client of FAH would produce occasional errors.  So I backed down to a happy 2.7ghz.  However, I then noticed that this motherboard was unlike any other I had dealt with...in order to get the RAM running in dual channel, the sticks needed to be side-by-side as opposed to staggered, as I am used to.  Once I put the RAM in their proper dual channel arrangement, the overclock was unstable.  I determined the cause of the instability was the motherboard handles SPD and RAM ratios differently when in dual channel versus single channel, so I had to work my way back up to 2.7ghz.  Along the way, I took note of how FAH output increased with the overclock.  My results are below. 

As you can see, FAH output increases lagged behind CPU clock speed increases by almost 50% with the HTT set to 2x and the RAM timings un-optimized.  Once reaching 2.7ghz, though, I raised HTT back up to 4x (times the FSB of 270 = HTT speed of 1080mhz) and set the RAM to an optimal speed/latency combination.  This tweaking pulled an additional 4.5% PPD improvement out of the PC (compared to the un-optimized single channel RAM setup), which translates to an additional 38 points per day!  That may not sound like much, but that's equivalent to the entire output of a lower-end CPU!  So, for a 35% CPU overclock, I managed to increase the PPD by almost 23%.  That isn't as successful as I would prefer, as I was hoping to see a 1:1 increase of PPD as CPU speed increased, but I suspect another factor is limiting the PPD output, like the relatively paltry 1MB of total L2 cache in this CPU, or the relatively low bandwidth provided by DDR running at about 450MHz (as compared to modern DDR2 or DDR3 running in excess of 1GHz).

Bottom line, there is some measurable advantage to tweaking your RAM and FSB (or HTT) settings when attempting to maximize folding @ home output.  These tests were done with a dual core AMD system, I would love to see some results with a quad core Intel system.  Any volunteers?  Hit us up in the forums!

---