I don't know if you guys are aware of this but there's currently a debate on what the best Tras (row active time) setting should be (from a performance standpoint). A lot of people have been saying that 11 is the optimum value (sounds too high to me), while others have suggested that following the formula Tras = Trp + Trcd + 2 produces the optimal result (haven't seen any proof behind that equation). So I decided to test it out for myself. The following is an excerpt from a thread I started over in the AMDMB forums.
Original thread at AMDMB forums:
http://www.amdforums.com/showthread.php?s=&threadid=240560
{BEGIN EXCERPT}
Some people have claimed that a Tras of 11 works better than other Tras values. Perhaps that's true on some systems, but it's definitely not true 100% of the time as some would lead you to believe. Below is my test report.
First of all, system specs:
- Abit NF7-S Rev 2.0
- Athlon XP 2500+ @ 2.2GHz (11 x 200) @ 1.7V
- OCZ PC3500 EL (512MB,
single channel) @ 200MHz @ 2.6V
I decided to test the difference between 11-3-3-2 and 5-3-3-2. After I had finished, I also decided to test out 3-3-3-2 to see if that made any difference. My testing procedure was as follows:
1) go into BIOS and select desired timings
2) boot into Windows (XP Professional)
3) start up SiSoftware Sandra Standard (version 2003.7.9.73)
4) run the Memory Bandwidth Benchmark portion of Sandra
5) record the value of
integer buffered memory bandwidth
6) repeat Steps 4 and 5 four times
Below are the results, expressed in megabytes per second (MB/s):
11-3-3-2
--------
2834
2839
2858
2860
2859
Average: 2850.0
5-3-3-2
-------
2833
2842
2858
2861
2860
Average: 2850.8
3-3-3-2
-------
2833
2841
2857
2860
2858
Average: 2849.8
UPDATE #1
8-3-3-2
-------
2835
2845
2857
2861
2861
Average: 2851.8
As you can see, the difference between the highest and the lowest average memory bandwidth is 1.0MB/s, which is
about one thirtieth (1/30) of one percent and easily within the margin of error. The difference between 11-3-3-2 and 5-3-3-2 would be more appropriately expressed in terms of
kilobytes per second than megabytes per second. Clearly then, there's effectively no difference between Tras settings of 11 and 5.
Bottom line: choose the Tras value that results in the most stable system. It sure isn't making any difference in performance.
UPDATE #1
As requested below, I ran my test procedure with a Tras set according to the following formula: Tras = Trp + Trcd + 2. Since I'm using Trp=3 and Trcd=3, that means a Tras of 8. Above is my result. As you can see, it is indeed slightly higher than the other Tras values. However, the difference is tiny: just 1.0MB/s between it and the next highest result, and just 2.0MB/s between it and the worst result. 2.0MB/s represents a difference of less than one tenth (1/10) of one percent. So this is hardly earth-shattering news, and it doesn't change the previous conclusion at all.
---------------------------
I've decided to do some additional testing, except that this time I varied CAS latency instead of Tras. Here are my results:
5-3-3-2
-------
2833
2842
2858
2861
2860
Average: 2850.8
5-3-3-2.5
--------
2780
2791
2804
2806
2806
Average: 2797.4
5-3-3-3
-------
2782
2787
2799
2806
2807
Average: 2796.2
The results are clear. There's about a 2% difference going from CL2 to CL2.5, but going from CL2.5 to CL3 produces almost no difference (
less than one twentieth (1/20) of one percent). Therefore, pushing up the RAM to 204MHz at CL2.5 would provide essentially the same performance as 200MHz at CL2. The conclusion is clear:
increasing RAM speed does more for performance than tightening the RAM timings. Don't be afraid to use CL2.5 or even CL3 (regardless of what the "hardcore" overclockers say about "loose" timings) if it means that you can increase the RAM speed by at least a few MHz.
---------------------------
Below are the results, as you requested. The RAM was set to 220MHz @ 2.8V, and everything else was untouched.
11-3-3-2
--------
3097
3105
3123
3124
3125
Average: 3114.8
5-3-3-2
-------
3092
3103
3121
3121
3121
Average: 3111.6
The difference is 3.2MB/s, or
slightly over one tenth (1/10) of one percent. Again, the difference is so tiny that it's easily within the margin of error. Thus we can conclude that even at high RAM speeds, my previous conclusion stands: there's essentially no performance difference between a Tras of 11 and a Tras of 5, so you should pick based on which value gives you the most stability.
{END EXCERPT}