You have a X6800? Lucky. Just for interest did you get a P965, i975X or other chipset? People say the P965 is better for overclocking, but the i975X gives better performance at a given clock speed.
In any case, I may have worded my response too strongly so hopefully I wasn't taken the wrong way. From your describtion I gather you were testing for decreases in performance scaling as you increased clock speed via multiplier at a given FSB speed. The linear scaling even on a paltry 400Mhz FSB is pleasantly surprising. I guess we have the large 4MB L2 cache and the superior prefetchers to thank for that since most of the FSB traffic is just pre-filling the L2 caches, something that is rather orderly and planned by the prefetches, rather than the FSB traffic being consumed by supplying information to satisfy pipeline stalls.
I was wondering if you could run some tests with the following configurations: 2.67GHz (133x20), 2.67GHz (167x16), 2.6GHz (200x13), 2.67GHz (267x10), and 2.67GHz (333*8) with as standard a memory speed as possible. I'm interested in this sort of configuration since it'll take into account the latency benefits of a faster FSB in addition to the bandwidth benefits. Personally, I don't think there will be much difference between a 1333MHz and 1067MHz FSB, but I'm thinking there may be a more noticable difference between 1067MHz and 800MHz and performance will drop off from there. Obviously, if you don't have time or aren't interested that is fine.
Your efforts are much appreciated and I'd certainly like to more of the linear scaling data that you've already obtained. Maybe you could make another chart like the one with SuperPi scaling?
Yep -- I originally slated myself up for an E6600, but then cancelled the order and put in for the X6800 for this experiment. People have berated me for the costs
but I am a fanatical more or less when I have a nagging question that I want to answer.
You are correct in the gist of the experiment and the reason I needed an X6800. Most OC reviews are overclocking via the system clock, this scales the FSB with the CPU speed and gives no indication if the FSB will choke the processor.
There are a few ways to get to this answer, hold the system clock fixed and drive the processor higher using multipliers. I have completed the 100 MHz system clock (400 MHz FSB) run using Cinebench 9.5, superpi at 1,2,4 and 8 M runs, Quake 4 1.2 with and without SMT, FEAR, Winrar, Everest memory, PCMark full, CPU, and memory suites. I am finishing up repeating some 200 MHz runs to verify the numbers.
Just quickly.
All apps are scaling linearly except winrar. Winrar is tracking 1:1 with measured memory bandwidth. This was expected based on how I am using it as a bench --- I am not crunching files, I am using it's random memory bench test. Using Perfmon it would appear the 4 meg cache is putting demands on the FSB of 5-25%, much less that I was expecting.
I am using a 975x MB (Asus P5W DH) initially I installed bios 1305, but ran into issues from a cold boot using a multiplier of 20, I then went to 1105 and was able to get data but ran into similar issues. I finally figured out what was going on so have now installed 1305 again and am repeating a few runs to make a consistent data set on 1305. Bios for most boards is still buggy.
Now, here is what is really fun about this chip... at 400 MHz FSB here are the idle full load temperatures ambient was 22 deg C, core temperatures measured with Core Temp.
Multiplier/Speed/Idle/Fullload
6/600/31/31
8/800/31/31
10/1000/31/32
12/1200/31/32
14/1400/31/32
16/1600/32/33
18/1800/32/34
20/2000/32/34
Phenomenal.
I plan to do a super post with a complete write up of all the data. I have collected screen dumps, super PI CRC numbers, and uploaded all my PC mark data to ORB so the data can be verified.
Sadly, I have been unable to really play with my new system because I have stayed at a clean install of windows and have only installed drivers, benchmark apps, and temperature monitors. Unfortunately, I was using Perfmon to look at cache miss rates, page reads/writes to work bandwidth. But Intel's Vtune software will measure FSB bandwidth directly, however when I installed it it changed window's system files to the point where now when the system gets to the welcome screen it reboots. So I need to rebuild the drive this weekend.
Finally, yes I will run your suggested experiments as soon as I can
...
Jack
Can you really load the system with generic apps and then do the games benchmarks. I don't like clean machine benchmarks. No one defrags and reboots before playing games or starting Cinebench.
No flames please.
Facebook
Twitter
Instagram