TCCD DDR600+ DFI nF4 Ultra/SLI guide

[kakaroto]

hey, i know this is mainly for Gskill....anyway i used this guide for my OCZ 3200 el rev2 plats and i can tell with CERTAINITY that they dont do 289 @ 2.5 3 3 7.
[i am running a 3500 winchester at 9x289/288, i am positive my CPU does 2600mhz)

The funny thing is ONE freaking mhz less (ie. 288) and i can run them finally 2.5 3 3 7 at 288. I dont know how many *hours* i tweaked and memtested.....but this ONE mhz was which almost drove me into insanity.

Eg. w/ 289 i was so on the edge that i passed 9 runs memtest #5 and got some errors on the 10th run ---> instable. My criteria is 15 runs of memtest #5 to consider them stable right now.

I can also say with certainity that *my* OCZ sticks run best at 2.9V....started out 2.7/2.8 but 2.9 is the golden sweet spot and 3.0 is the same...so 2.9V is it, no more, no less. memtest #5 is a nice indicator for needed voltage.

One finding:

I think my efforts (and also the recommendations to IF ANY POSSIBLE run 2.5 3 3 7 instead of 2.5 4 3 7) are more "theoretical" than doing anything in real life.

My mem bandwidth (after tweaking) in memtest is always around 3176MB/s...and after i finally managed to run the sticks 2.5 3 3 7 (with 9mhz less on the CPU since i changed from 9x289 to 9x288 but with TRCD@3) i did my benchies.

Eg. Farcry 800x600 Hardware OC benchmark (which is good for testing latency/mem bandwidth impact etc.) has the same results as i have with the former TRCD @4 - and everest/memtest etc. is all about the same.

I cannot confirm that TRCD at 3 is some magic goal to achieve for great improvement...at least i dont see ANYTHING which would make me conclude that in any benchmark results !
In other words:
If you run 2.5 4 3 7 in the 290-3xx HTT range...and you get ~3200MB/S in memtest then you might be well off and can save yourself the time to go to 2.5 3 3 7 since it doesnt do ANYTHING. Zero !

Feel free to prove me wrong by showing me a benchmark where i can see a huge impact by changing TRCD 4-->3. At least not in my case and at speeds in the HTT range 290-3xx.



NOTES:

Beware of "async latency"....eg. set it to 9ns and it passes memtest.....but i get bad crashes in windows. So leave at 8ns or 7ns.

Beware of every memory benchmark in windows...especially everest since the results can change GREATLY - depending when you run it.
GREAT results right after boot - do some stuff, run everest again.....and the values hit rock bottom. This is hardly reliable and usable for comparison w/ other people.
IMHO the only halfway reliable measurement for bandwidth is from within memtest.

First, thank you for sharing your experience!

The difference between Trcd 4 to 3 can only be seen in benchmarks.

If you test Sandra 2004/2005 you can see some bandwidth loss when running trcd 4 instead of 3.

Here a sample with Pifast calculating 1M pi:



MEMTEST86++ bandwidth is not accurate. I have experienced heavy bandwidth difference when changing Trcd, Tref, Twr and Trrd, sometimes 100-200MB difference.
The bandwidth is also related to your processor speed, not constant like on the S754 platform.