Ok, there is still more work to be done here ... but I have some initial results for you.
Rig Setup:
Asus Rampage II Extreme (BIOS 1306), Core i7 965, 12 Gigs of DDR3-1333 (C9), 2 GTX295s in SLI, 1 9600GT (drives second monitors), 6 HDs (3 2x1 Raid0), Thermalright 120 1366 edition, TT 1200W PSU, Vista Ultimate 64 SP2 with oodles of crud loaded at startup (you should see my tray, very messy not good).
Tests:
I ran tests at 2.66, 2.80, 2.93 GHz (using the core multiplier) -- as such, Asus BIOS disables turbo mode, screen shots are located in the gallery link attached. Each run was observed via CPUID to run at the set multiplier. Due to running at DDR3-1333 the NB frequency was set at 2.66 GHz (has to be 2x the effective mem speed, this I don't think will match Anand's setup). I ran 2 tests at each clock speed, one with HT on and one with HT off. In most cases I ran the 1CPU test twice (not all, but most cases). So most test ran 1CPU/XCPU/1CPU and in cases where I ran twice I report the second run value below. EDIT NOTE -- I turned off the second monitor such that when I did screen captures I did not get 3840x1200 images
...
I report the data below as such:
CPU Clock/NB Clock/HT on or off -- 1CPU/XCPU, e.g. 2.66/2.66/On is 2.66 GHz core clock, 2.66 GHz NB clock, and HT is on.
2.66/2.66/Off -- 3709/13165
2.66/2.66/On--- 3501/14735
2.80/2.66/Off -- 3907/13386
2.80/2.66/On -- 3620/15453
2.93/2.66/Off -- 4124/14278
2.93/2.66/On -- 3805/15953
Link to screens (note image names follow same convention as above) -- please check me for typo's:
http://forums.xcpus.com/gallery/v/Ju...sc_001/i7CB10/
So it is indeed possible that Anand's i7 920 turbo'ed by +2, though frankly I have never observed this on the 965 (of course I am using a different stepping than he is). Also, my NB frequency is not likely matching Anand's since I ran at DDR3-1333 and the NB must be 2x or higher the RAM speed. Finally -- Notice the number of processes running -- > 100 (the rig is currently way bloated and it was to much troubles for this initial run to go through and disable all the unnecessary services).
In my experience with i7, I can get +2 turbo on single or dual threaded apps when the app schedules itself exclusive to one core (or two as the case may be). CB10 uses the Vista scheduler (which just sux for the most part), as a result threads are distributed over all cores (HT or no HT), which I believe (not entirely sure) is enough that the turbo algorithm does not kick up into the higher mode.
Nonetheless, his Lynnfield data is still discrepant on this test in my opinion, so I will see what I can do to find some time to flesh this out more. Including a fresh install of Vista without all the bloat.
I do plan to work on this some more, and study just how much impact the NB speed would have -- and with a clean build (no bloat) but it will take a few days.
Jack
Bookmarks