CL3 vs CL4 vs CL5 + 1T vs 2T

INTRODUCTION
I spent long hours up last night finding max frequency at each standard CAS Latency.
I was freezing my behind off in front of an Air Conditioner to keep
case cool. I drank loads of scotch and coffee to stay awake.
I also corrupted every second file in windows, RAM has a tendancy to do this
when it is undervolted for the speed you are trying to boot at.
The reason for this testing Frequency VS Cas Latency, Results found
in bandwidth (read, write and copy) and also in Latency (ns)

I was also testing how far I could take 1t Command Rate.
Interestingly 1t had little effect on total RAM OverClock for CAS3,
absolutely no effect on CAS4 (maybe a few Mhz but I did not want to change FSB).
However it did effect how far I was able to push CAS5.

TEST METHOD
All testing done with AC blowing into side of case (freezing cold) and vMEM
at 2.425v BIOS for about 2.47 REAL volts.
I used Everest and SPI (XS mod) to test results. One thing should be noted,
regardless of whether I had Max speed at CAS3, CAS4 or CAS5, all of my
times on SuperPi were nearly identical, No more than 0.05 second's of
difference, so I did not keep results. SuperPi1m times were all around 14.141s.
I Did not have any fancy hardware to use, so I just used my rig.
CPU: C2D E6600 @ 3.6ghz (1604 FSB)
Mobo: Asus P5N32-E (680i)
Ram: OCZ Sli ready PC8500 CL5 (DDR 1066)
GFX: 8800GTX
PSU: Thermaltake TP 850w

Each test was run twice, I took the screens of second run.


Now onto the results.


Tight Timing's
Max Frequency @ CAS3 and CR1t
http://img126.imageshack.us/img126/3994/cl31tot8.jpg

Max Frequency @ CAS3 and CR2t
http://img45.imageshack.us/img45/9701/cl32tkr1.jpg


Medium Timing's
Max Frequency @ CAS4 and CR1t
http://img299.imageshack.us/img299/5830/cl41tva7.jpg

Max Frequency @ CAS4 and CR2t
http://img113.imageshack.us/img113/2706/cl42tdi3.jpg


loose Timing's
Max Frequency @ CAS5 and CR1t (the 1T CR limited me badly here.)
http://img113.imageshack.us/img113/1533/cl51tey2.jpg

Max Frequency @ CAS5 and CR2t
http://img159.imageshack.us/img159/876/cl52tjh3.jpg


CONCLUSION
It is easy to sit down and do the mathamatics to work out
which combinations of timings vs frequency will benifit you the most,
but there is very little I was able to find where testing was performed to
back the theory up.

Here is a field example for anyone who can not decide whether they want
to use stock timing's and high frequency, or whether they want to set up
tighter timing's.

It was a BIG surprise to find that CAS5 max frequency had such a low latency but CAS3
max frequency had higher latency but more bandwidth. Quite the opposite of what I expected.

So in the end which is the best? Well you can decide, here are my results.

Good work.
I have a 680i , and that is pretty much what I get too.
We just have to remember, that this is only valid, for 680i... On recent x38, the higher MHz DDR2 clock with CL5 will perform better then the lower MHz DDR2 clock with CL3. At least, this is what I have seen here on XS.

[]'s
Simps

Quote:

---

Originally Posted by Simps

Good work.
I have a 680i , and that is pretty much what I get too.
We just have to remember, that this is only valid, for 680i... On recent x38, the higher MHz DDR2 clock with CL5 will perform better then the lower MHz DDR2 clock with CL3. At least, this is what I have seen here on XS.

[]'s
Simps

---

O yer, true, I mentioned that I was using 680i in test.

I should also say that at lower frequency the 1T certainly helps with both latency and bandwidth.

But it is a rough guide for any one to use how they choose.

Yeah!
This stuff is from the user "Pyro"
Looks good for 450Mhz DDR2, doesn't it?

http://img404.imageshack.us/img404/5...033341thm3.jpg

[]'s
Simps

Yer thats amazing, I cant get anything like that, probably not even get a sniff of results like that. But I am just using my every day rig, with my nice and cheep OCZ sli RAM lol



ACTUALLY I can get bandwidth like that when I use 450 FSB and 8 Multi but then I end up with over 54ns of latency lol.



NEXT Time I shall use 450 FSB for strap change and see if i can pull down latency also. I will do all the same tests, but with 450 FSB not 400.

I have a Striker Extreme, and I can't get close to that either... Very impressive stuff.

[]'s
Simps

If some one want's to send me some ballistix that I can use at 2.6v I would try lol. But I still doubt I could do that. Maybe though.

The secret is, he was @ 450FSB / 450MHz DDR2 1:1 Linked / SYNC.
And he was running CL3 / 1T, with sub-times optimized on memset, and a very good "PL" (performance level set on memset).

Of course, he has good rams to handle it :)

[]'s
Simps

Quote:

---

Originally Posted by Simps

The secret is, he was @ 450FSB / 450MHz DDR2 1:1 Linked / SYNC.
And he was running CL3 / 1T, with sub-times optimized on memset, and a very good "PL" (performance level set on memset).

Of course, he has good rams to handle it :)

[]'s
Simps

---

Yer I doubt my RAM can handle anything like that.

I also did not use mem set, all tests are set from bios and have to boot, run spi 1m, then this test twice and take screen. Did not bother to test stability past that.

Next week I might do all tests again with 450mhz FSB to report on strap change. I will also try my ass off to get a score like Pyro's lol...

I would love to know what RAM, Voltage and timings/sub-timings he has used!

CR 1T @ Striker Extreme was a bad experience, I had stability problems even at low clocks between 450-500Mhz.
Sry, I don't have Everest screen, but this one I want to share. :yepp:

OCZ 2*1GiB DDR2-800 Crossfire @ 2,30V (real)

http://www.img4free.com/images/Hyper...5hyperhorn.jpg

tRTW=2 tRDRD=3 tWRWR=2
DDR2-925 3-3-3-2 is rare within the guarantee. ;)

Quote:

---

Originally Posted by The0men

Yer I doubt my RAM can handle anything like that.

I also did not use mem set, all tests are set from bios and have to boot, run spi 1m, then this test twice and take screen. Did not bother to test stability past that.

---

any chance you have the sp1m times written down to show the performance difference?

and why are you at 42 - 46 tRFC?

anyway, nice writeup there.

Quote:

---

Originally Posted by naokaji

any chance you have the sp1m times written down to show the performance difference?

and why are you at 42 - 46 tRFC?

anyway, nice writeup there.

---

Have high tRFC purely so it was not going to be a limit for higher frequency, going into 30 odd and I can not get as high frequency.
If I leave it auto it would have gone to 50 and beyond, which is only reason why I set it manually.

Yes I know the SPi 1m times, I dont have screen but I can tell you.

Fastest was with timings Cas 5-5-5 2t (14.141s) and high frequency. But also Cas4-4-4 1t did the same. Cas3 best was also with 1t but time was 14.17xs I cannot remember Exactly. But if you would like me to do a run at a couple of the specified reasults I can, as I will probably set up and do more runs. When I get my timing's optomised, and I learn how to tweak for SPI I will try and get very good SPI 1m with 3.6ghz.

Quote:

---

Originally Posted by Hyperhorn

CR 1T @ Striker Extreme was a bad experience, I had stability problems even at low clocks between 450-500Mhz.
Sry, I don't have Everest screen, but this one I want to share. :yepp:

OCZ 2*1GiB DDR2-800 Crossfire @ 2,30V (real)

http://www.img4free.com/images/Hyper...5hyperhorn.jpg

tRTW=2 tRDRD=3 tWRWR=2
DDR2-925 3-3-3-2 is rare within the guarantee. ;)

---

Crikey mate thats terrific. Gee I wish I could get that out of my sticks. You should do an everest with something like that.

Edit: what is your northbridge voltage and cooling?
To be able to get this I could not use stock heat sink, and I use 1.5v Northbridge so i have lots of give there.

Well the golden points on nVidia based boards are high clocks with 1T or ultra low timings with around 800Mhz , on Intel based boards and cpus it's best to keep it lets say 1000Mhz 5-5-5-15 2T (very mild setting) or around 900 - 1000Mhz 4-4-4-12 also 2T ( because not all Intel based boards have 1T selectable in bios)
Current Core2 architecture likes high bandwidth , the thing is that in older DX8 and DX9 based game titles the low timings are better but the new games like medium timings and higher bandwidth 1T or 2T is not that important. But that's is all generalized statement , it all depends on your personal experience : ]

@Hyperhorn - great timings from these OCZ , are they 800Mhz 4-4-4-15 1T EPP ? i have EPP edition but for now i use the EPP profile only with higher clock, i must try to push them @ CL3 ...... fascinating :f

@The0men - you have E6600 @ 3.6Ghz @ that speed i have super pi around 13.800ish or something like 14 , but once you go over the 3.8Ghz there will probably be bandwidth "hunger" symptoms ;] or it's just the cooling vs. limitations of the architecture of the cpu. It takes no genious that once you get past the 3.8 - 3.9Ghz it doesn't scale with the time/speed of cpu scores , i mean with sub 4Ghz every 100Mhz on cpu gives around 1sec time less to compute , after the 3.9 it takes 300 or 400Mhz to do that ;]

nah.. no need to run specific tests for me, was just looking for what scales better, latency or clocks... seems like high clocks would be the way to go for sp1m from your numbers. (cant wait till my guaranteed d9 team ram arrives... arggh i hate the speed of royal mail).

SuperPI runs must exceed L2 cache of cpu to get memory work with it.
So even for new Quads@12MB cache you have to run at least 16M to
see any differences. On the other hand everest can show the differences.
It is pointless to change strap(performance level) to see what you got.
Keep it the same so as to have a clear idea only for memory timings.
Of course different chipsets have different tweakings.
If i have time i will try what i have told although i find it difficult on this mobo
to run 3-3-3-x (never tested so far).
Also alpha timings are important.
My runs will be as:
1)4-4-4-9-24-6-2-2-3-1T
2)5-5-5-13-30-7-3-3-3-1T
There is another middle timing that my memories can handle with medium voltage
5-4-4-9-24-6-2-2-3-1T (only CAS is changed)

the problem is that tight timings require a rather big drop in clocks while you cna ramp clocks up pretty far with mediocre timings.... thats why i came up with high clocks being better than low latency.

I have also found I have no worries hitting ddr 1333 with CAS 5, This theotetically speaking is still as fast (if not faster) than what I have been achieving. So I may try CAS6, I have a feeling that it will not act as poorly as people expect.

Damn I wish I had some Gem sticks.

@naokaji-
There has always been alot of debate about which is the fastest. This is why I made this thread. It is just to show from my experience what I have found. But SPi Is not a good example, because all of my results were within a margine of error (I might move my mouse during the test, or windows might still be loading.) And I do not know much about spi but what Spyrus said seems to make sense. Everest is good because it is a quick way to get readings. But your reading can vary alot from one test to the next without a single setting changed. So it is only for a guide.

Anyways I would still like to preform a couple more test with high FSB, I am sure my latency will go up, but so will my bandwidth dramatically increase.

I am also going to try a couple of those really tight timing runs. But my RAM just isn't quite up there. It's 650mhz club member, but not 700.

very nice results, thanks for posting them here :toast:
cas6... forget about it... the max clockspeed difference from cas5 to cas6 is very very small

Quote:

---

Originally Posted by Hyperhorn

CR 1T @ Striker Extreme was a bad experience, I had stability problems even at low clocks between 450-500Mhz.
Sry, I don't have Everest screen, but this one I want to share. :yepp:

OCZ 2*1GiB DDR2-800 Crossfire @ 2,30V (real)

http://www.img4free.com/images/Hyper...5hyperhorn.jpg

tRTW=2 tRDRD=3 tWRWR=2
DDR2-925 3-3-3-2 is rare within the guarantee. ;)

---

Hyperhorn, up to now I did not find the time to talk to you on PCGH Extreme about this CL3 result.

You used a very old version of MemSet when doing it, have you tried to change the CAS latency in Windows? Can you do this when booting?
Because I think you would need much higher volts to do this.

Quote:

---

Originally Posted by saaya

very nice results, thanks for posting them here :toast:
cas6... forget about it... the max clockspeed difference from cas5 to cas6 is very very small

---

Thanks. Hopefully I can try for more results during late next week. Maybe i'll also have a clean install of windows.. This one gone to the shi(f)tter something cronic. After its last corruption for the night (kids dont try this at home) all my fonts are italic no matter what i do.. all my start menu is gone. Well infact, lots of things are gone lol. Not going to bother reformatting until I have to though lol

Yes I cannot see a good improvement for CAS6. I guess this is why they don't use it in the mainstream market.

Quote:

---

Originally Posted by Simps

The secret is, he was @ 450FSB / 450MHz DDR2 1:1 Linked / SYNC.
And he was running CL3 / 1T, with sub-times optimized on memset, and a very good "PL" (performance level set on memset).

Of course, he has good rams to handle it :)

[]'s
Simps

---

Actually Memset was not used at all ;)
It was just a very quick benchmark (had a couple o mins free), everything was set in the BIOS and booted straight into Windows.

The Everest run was all I did before setting it back again. (and stability was not tested at all)
If I remember correctly, it just took a lot of NB volts mainly...
I *think* the RAM was set to somewhere around 2.5-2.55V (didn't take notes :rolleyes: )

Basically, for bandwidth, I found that 450MHz for FSB was optimal, 1MHz higher and the bandwidth and latency drop as the board changes straps.

Quote:

---

Originally Posted by Pyr0

Actually Memset was not used at all ;)
It was just a very quick benchmark (had a couple o mins free), everything was set in the BIOS and booted straight into Windows.

The Everest run was all I did before setting it back again. (and stability was not tested at all)
If I remember correctly, it just took a lot of NB volts mainly...
I *think* the RAM was set to somewhere around 2.5-2.55V (didn't take notes :rolleyes: )

Basically, for bandwidth, I found that 450MHz for FSB was optimal, 1MHz higher and the bandwidth and latency drop as the board changes straps.

---

Thank you for your info. Yes I may need more volts on NB if I am to even get half way. I was running 1.5v already I think.

memory latencies don't affect intel systems as much as they do K8 or K10 systems, right?