nForce2 and Memory - odd observations

[SteveI]

To experts of the NF2 chipset, I apologize if this is obvious to you... I recently discovered a few odd tidbits in my random experiments, and wanted to share. I am by all means no expert on this matter, and would appreciate some insight and interpretation... I also welcome disagreement or conclusions.

First observation, NF2 and crucial pc2700 don't seem to mix. I noticed this a while back and thought is odd. I had 2 sticks of Crucial 256meg PC2700. Both sticks ran fine w/ tight timings in the 180 to 193mhz range on KT333 mobos. I tried to use one stick in an 8RGA+, and the stick couldn't do default timings, even w/ 2.7v and at 160mhz.. It seems to top out at 155mhz at slow timings. Due to system issues, I tried the other stick today on my NF7s, with only slightly better results. I get 161mhz at slow timings and 2.7v. Any higher yields Memtest errors. Both mobos handle high fsb fine with other branded sticks.

Second observation. CAS has minimal impact on performance... but Row Precharge Delay and RAStoCAS delay are much bigger. To support this, here are some numbers.. I ran 159mhz for all 3 runs and 2.6v


run--CAS-Row pre.-RAStoCAS-RowActive-Score
----- lat - delay --- delay ----- delay
1 -- 2.0 -- 2 -------- 2 ------- 11 ------- 758
2 -- 2.0 -- 3 -------- 3 ------- 11 ------- 652
3 -- 2.5 -- 2 -------- 2 ------- 11 ------- 746

[edit: sorry if the table is hard to read, I tried to make it more readable.]
Score refers to the raw score given by memtest. I don't have time to run sisoft, and the results w/ sisoft seem too inconsistent anyway. For these runs CPU interface was turned off. Run 3 is an eye opener for me. I do not know if this holds true in dual channel mode, but once my replacement sticks come in, I will do more testing. ATM, I find this quite interesting.

CAS latency used to be the "end all be all" when using via chipsets, but now seem almost insignificant. At least for nf2 systems, CAS will be the first thing I sacrifice to gain stability.

I realize this is not an in depth analysis, but seems to beg for further "look sees".
Anyone else see similar results?

I've heard all systems are unique and results will vary. If anyone has a different mobo, and get opposite results, please let me know. When I have time, I can try to duplicate this little experiment with my son's 8rga+.

[JAWS]

SteveI you are absolutely correct. I bet you I tried every memory combination, followed by a Sandra benchmark and Row Precharge Delay has a great effect on performance. 11 is by far the best setting, then maybe 7.

[SteveI]

Originally Posted by JAWS

SteveI you are absolutely correct. I bet you I tried every memory combination, followed by a Sandra benchmark and Row Precharge Delay has a great effect on performance. 11 is by far the best setting, then maybe 7.

Yup... 11 is better than any other setting 5-13, regardless of other timings including cas. Odd how that works out.

My gig o' memory made it to my front door today.... My wife wasn't home so it's back at the FedEx depot... argh!! I'm dying running at 256megs @ 159mhz.. I can't wait to crank it back up again. (You never appreciate what you have until it's gone.)

[Holst]

Strange results.

I think they may be more down to the way memtest produces its score than anything.

What ive noticed is that CAS wont often make a difference in memory benches where your just testing throughput.. then its the mhz that counts.
This is definatley true of sandra.

But in 3dmark and others timings easpecially cas are very important.

11,1,1,2 is the fastest, although 11,2,2,2, is the best you can run (bar a few people who can get 11,2,1,2 at low fsb)

But the most important thing is CPU interface, makes a huge difference to my maximum bandwith and 3dmark score.

[SteveI]

Thanks Holst... that's more to think about.
I haven't done any 3dmark testing with this memory. Once my PC is back to a gig. I'll mess around with it some more.

[Holst]

I find that just running low detail dragothic in 2001 is a pretty good and fast test for memory bandwith.

[Áedán]

Given that the nForce 2 chipset does some kind of speculative lookahead for memory fetches, the CAS effect could be mitigated by that.

Basically, the nForce2 chipset seems to guess what data's going to be required next, and in the case of memory benchmarks that read linearly, probably guesses correctly 99.9% of the time. Hence, it's already fetching the data before the processor asks for it.

Drop back to real world examples, and I suspect that the boost that provides drops away rapidly. Still better than nothing however!

Basically, most of the memory benchmarking tools aren't really much good for benchmarking the nForce 2 chipset (and probably others too). The chipset does too good a job of hiding various latencies.

[JAWS]

I thought about this thread while reading this one:

http://www.xtremesystems.org/forums/...threadid=22071

[SteveI]

Here's another observation I find odd... I'm now running dual mode:

@217 fsb 2.5-3-3-11 w/ 10x, memtest scores 1571
at 11x the score increases to 1581.

@217 fsb 2.5-2-2-11 w/ 10x, memtest scores 1696
at 11x the score decreases to 1666!!!

Can anyone make sense of this for me???
When I realized these results, I went back and duplicated them many times, making sure nothing else changed.

[SPQQKY]

The only memory benchmark I have used that truly shows the benefits of nForce2 is PCMark2002. I have the same problem SteveI, and I believe that perhaps your cpu is reaching a point of instablity at 217 x 11 which will effect you memory bandwidth. Unfortunately, my memory will not run 11-2-2-2.5 at anything over 166MHz even though it is TwinMos pc3700. It will run all the way up to 240 at 11-3-3-2.5. The Trcd seems to be stuck at 3 for this RAM.