IMO, good 1:1 high timing vs 5:4 low timing thread

[aggressor]

http://www.hardforum.com/showthread.php?s=&postid=1025224557#post1025224557

Pretty good info, not really what I expected though

 

[oldfart]

A few more real world benches would have helped. I've posted about this like a Gazillion times now, but people still dont seem to get it.

1) DDR speed AND timings matter for actual system performance.
2) 1:1 with high latency ram is no faster (sometimes is SLOWER) than running 5:4 with low latency timings. This has been proven many times over in actual application benching.
3) SiSoft mem benches are a joke. They DO NOT translate into real world performance. I've see plenty of benches showing huge gains in SiSoft that show no increase or even a DECREASE in performance when real world benches are run. The only thing that bench is good for is for mfgrs to use to sell ram to people who dont know any better.
4) CPU speed matters MUCH more than any mem speed/timing tweak. Set your overclock up for the best CPU speed, then get whatever you can out of the ram.

In reality, 1:1 slow timings and 5:4 fast timings give ~ the same performance. This idea that you have to run 1:1 and PC3700 - 4000 for good performance is a fallacy.

 

[aggressor]

Look down a bit, a guy posted game benchmarks

 

[oldfart]

I can point you to many reviews and posts like that one that show the same thing. SiSoft scores mean jack. I've also tested this myself. Still, people will keep buying the PC3700/4000 ram because of the SiSoft score.


People who think Synthetic scores like SiSoft/Aida matter

---------------------LINE IN THE SAND--------------------

People who care about actual system performance

 

[DerwenArtos12]

This has been annoying me forever now. People just because I say that 1:1 is better does not mean that you should get cas3 ram that is 500mhz. Crank your multiplier up and your fsb up that you have some change at hopefully getting 1:1 but having a 2500 at 10x200 with 1:1 is not optimal. you should be getting 220x11 with 1:1 on a good cas2 ram for perfomance!

 

[DerwenArtos12]

oh and i did find a use for sisoft. the scores can be useful when comparing stock speeds to oced speeds. so long as you are comparing your results to others of your results. then it is not completely accurate in performance gains but it is pretty good.

 

[THUGSROOK]

and dont forget me!
my system puts a huge wrench in the debate over 1:1 slow cas vs 5:4 fast cas ~ cause i have 1:1 fast cas!

 

[boshuter]

Originally posted by: THUGSROOK
and dont forget me!
my system puts a huge wrench in the debate over 1:1 slow cas vs 5:4 fast cas ~ cause i have 1:1 fast cas!

Just had to rub it in didn't you...

 

[oldfart]

Originally posted by: THUGSROOK
and dont forget me!
my system puts a huge wrench in the debate over 1:1 slow cas vs 5:4 fast cas ~ cause i have 1:1 fast cas!

So true. Only thing is, you dont have Hyperthreading which can give a significant performance boost in the right applications. This improvement can be greater than any memory option we are talking about here. To get everything: Hyperthreading, 1:1 ratio, fast timings, you would have to use a 3.06 CPU. The cost of it is way too high though.

 

[pallmall]

Originally posted by: oldfart
A few more real world benches would have helped. I've posted about this like a Gazillion times now, but people still dont seem to get it.

1) DDR speed AND timings matter for actual system performance.
2) 1:1 with high latency ram is no faster (sometimes is SLOWER) than running 5:4 with low latency timings. This has been proven many times over in actual application benching.
3) SiSoft mem benches are a joke. They DO NOT translate into real world performance. I've see plenty of benches showing huge gains in SiSoft that show no increase or even a DECREASE in performance when real world benches are run. The only thing that bench is good for is for mfgrs to use to sell ram to people who dont know any better.
4) CPU speed matters MUCH more than any mem speed/timing tweak. Set your overclock up for the best CPU speed, then get whatever you can out of the ram.

In reality, 1:1 slow timings and 5:4 fast timings give ~ the same performance. This idea that you have to run 1:1 and PC3700 - 4000 for good performance is a fallacy.

Then what about 1:1 with lower fsb, higher multi and fast mem timings?
What you guys think the advantage of getting higher fsb given the same clock speed(fsb*multi)?
If the effect of sync mem(1:1) is better than higher fsb, I would rather go fsb-10, multi+1, sync & fast memtimings.

 

[THUGSROOK]

there is no "perfect system" at the moment.
no matter which P4 you buy there will be some sort of over overclocking "comprimize".
i would just rather error on the cheaper side

if i did (for some reason) decide to go with an 800fsb C chip ~ id be planning on using a 5:4 ratio. ive seen enough comparisons (including my own irrelivant ones) to show that cas settings are faster then the high $ ram. (and cheaper)

my results only show how well 1:1 could be on cheap ram if the right chips were available (which they arent unless you want to buy a high $ 3.2C)

 

[orion7144]

I am running 275FSB 1:1 (2.4C) with OCZ PC3700 Gold memory. If anyone has looked at the benchmarks I put up several times comparing the 2.4C w/the Gold vs the OCZ PC3500 running 5:4, at the same memory clock speeds there was NO difference in going from CAS2 to CAS3. The benchmarks also proved that a higher FSB was alot better than a higher memory bus. Hence if you can get a higher FSB using the 5:4 divider was alot faster than running 1:1 at a lower FSB.

Thugs had posted on a different forum about how maybe the CAS settings were backwords on the 865/875 Abit boards. This maybe true or it may just not mater on the faster boards as I noticed virtually no difference in memory bechmarks with CAS2/2.5 and 3.

I do know that OCZ has gained a new lifetime customer.

 

[pallmall]

Originally posted by: orion7144

higher FSB was alot better than a higher memory bus. Hence if you can get a higher FSB using the 5:4 divider was alot faster than running 1:1 at a lower FSB.

I am not quite sure which performance you are talking about : sandra mem bandwidth or superPi(cpu+memory).

 

[orion7144]

Originally posted by: pallmall

Originally posted by: orion7144

higher FSB was alot better than a higher memory bus. Hence if you can get a higher FSB using the 5:4 divider was alot faster than running 1:1 at a lower FSB.

I am not quite sure which performance you are talking about : sandra mem bandwidth or superPi(cpu+memory).

Cpu + Memory. The memory bandwidth did not change enough to justify lowering the FSB. Ofcourse I don't have to wory about it now since I am running at 275 1:1 but when I was running 233 1:1 with PC3500 changing to the 5:4 divider and running 275 was alot faster and more responsive.

 

[oldfart]

Originally posted by: orion7144

Originally posted by: pallmall

Originally posted by: orion7144

higher FSB was alot better than a higher memory bus. Hence if you can get a higher FSB using the 5:4 divider was alot faster than running 1:1 at a lower FSB.

I am not quite sure which performance you are talking about : sandra mem bandwidth or superPi(cpu+memory).

Cpu + Memory. The memory bandwidth did not change enough to justify lowering the FSB. Ofcourse I don't have to wory about it now since I am running at 275 1:1 but when I was running 233 1:1 with PC3500 changing to the 5:4 divider and running 275 was alot faster and more responsive.

See my point #4 on my first post. CPU speed matters MUCH more than mem speed. Dont give up CPU speed for memory speed.

I'm not sure how you did your testing, but memory timings certainly DO matter. My testing, Thugs, countless others have shown that to be true.

 

[pallmall]

Originally posted by: oldfart
[See my point #4 on my first post. CPU speed matters MUCH more than mem speed. Dont give up CPU speed for memory speed.

I'm not sure how you did your testing, but memory timings certainly DO matter. My testing, Thugs, countless others have shown that to be true.

I agree cpu speed(fsb * multi) is the number one factor for the performance(cpu+memory).

My question is what is the second most important factor for the performance :
(fsb or memtimings) combined with mem(sync or async)?

Complicated ? In order to compare the factors correctly, we need to restrict a factor stable.
The cpu clock(fsb * multi) is mostly restricted by it's overclockability.
To make the cpu clock stable, we need to adjust the fsb by the multi change or vice versa.

1. high fsb(low multi), mem sync, slow memtimings
2. high fsb(low multi), mem async(4/5), fast memtimings
3. low fsb(higher multi), mem sync, fast memtimings
4. low fsb(higher multi), mem async(5/4), slow memtimings

Except #4 looking stupid #1,2,3 could be good candidates for comparison.
Any Idea or results?

 

[oldfart]

1:1 with slow timings and 5:4 with fast timings are really close. You will see some applications prefer one or the other.

 

[pallmall]

Originally posted by: oldfart
1:1 with slow timings and 5:4 with fast timings are really close. You will see some applications prefer one or the other.

Thanks,
One more thing to be added to the restriction : benchmark applications.
What about restricting it to SuperPi or 3dmark2001 or both?
It is really hard to get the information covering all major variables.
They tend to emphasize only the partial variables making it very hard to compare correctly.

 

[oldfart]

I always use benches that are real world when testing. Not everyone uses the same applications. Game, Video/audio encoding, archiving are a good mix.

 

[pallmall]

It is making the comparison impossible.

 

[oldfart]

Dont know what you are looking for. Its not perfectly cut - n - dry. There have seen many forum type tests of this and several reviews on the web. It is as I posted. They are similar in performance. Usually, low latency will win out over high DDR speed. There are too many variables to just give a black and white answer. There are different CPU/mobo/FSB/ram/brand/ratio/timing/PAT/etc/etc options. Generally, is all I can give you. 1:1 high latency and 5:4 low latency give similar results. To go further, you would need an exact CPU/FSB/mobo/ram brand setup to compare.

 

[pallmall]

As far as I know, all the scientific tests can be proved by controlling the variables.
But I cannot say they are pursuing the b/w theory.
Declaring something is easy, but being understood and accepted is another thing.
Even though I feel something is right, it is still the assumption not the truth.
Just say it is assumption or prove it with real data and model.

 

[oldfart]

The many benches that have been done to prove it are not enough for you? The data is readily available. Why not do a few of them yourself?

 

[pallmall]

I thought you have noticed it already.
They do not conclude the same. Mainly by different conditions or purposes.
It is very important to check the conditions under which the tests were performed.
The conclusion maybe only useful under the same condition.
But it is too dangerous and impossible to generalize them(under any conditions) into a theory.
I did not ask your opinion on everything, which is impossible to be answered.
I just wanted to provide some different aspects for a wider consideration.
If you can give your opinion for the question, it is very good and appreciated.
Please don't assume I have done nothing while people become elite by posts.

 

[oldfart]

Dude, I dont understand you at all. You are making WAY TOO MUCH OF THIS. What are you some scientific method test freak or something??? .

The tests have been conclusive. 1:1 ratio, high DDR speed, high latency performs ~ the same as 5:4 ratio, lower DDR speed, ow latency with a slight edge going to the latter. That is what all the tests all over the net have shown. Each has shown very similar results. I dont see what is so difficult to understand here.

Please excuse me for not testing each of the 10's of Thousands of CPU/mobo/ram/FSB/timing options for you. You are welcome to do so yourself. No one gets paid anything here. We just share our knowledge of what we have tested and seen around. You needn't micro-analyze this any longer.