Here Thlnk3r, this testing is for you and Joe:
I thought I would investigate more-fully the ganged versus unganged memory controller option. I took Everest Bandwidth tests 7 times under each setting. The lowest and highest values were discarded, the other five were averaged.
I also decided to try some applications. I used the fastest running program I have (about 25s per run) and did similar. For the single SuperPi application, it is easy: run, record; run, record; et cetera. For the SuperPi4x, it was a bit more complicated. I opened SuperPi from a folder labeled 0, and set affinity to Core 0 solely. A SuperPi from a folder labled 1, and set affinity to Core 1 solely. The same for Cores 2/3. Then I told all four to calculate 1M, but did not hit the "okay" to start the run. I stacked all the buttons up, and then rapid-fired the mouse button, causing all four cores to be operating within milliseconds of each other. The four cores were then averaged.
A single instance of SuperPi benefits by 0.5 % by using ganged.
Four instances of SuperPi benefit by 1.7% by using unganged.
Memory read benefits by 12.5% by using ganged.
Memory write benefits by 0.3% by using unganged.
Memory copy benefits by 8.3% by using unganged.
Memory latency benefits by 0.8% by using ganged.
I might try some 3DMark06, but given the minute differences in bandwidth, I do not see any real changes.
Also in the agenda: Another 2 MHz on the HTT and/or lower RAM timings.
And for the record, the address in my prior post is "C:\\Users\\Keith\\Docume~1\\Stress~1\\SuperPi ." The CODE tag seems to delete that character.