The CAS latency myth

Resources, guides, discussions for modding your system.

The CAS latency myth

Postby Valour » Wed May 14, 2003 2:52 pm

Recently I set out to determine what the real difference was in performance between CAS2 and CAS3 RAM. I wrote an article on it here but more evidence suggesting that CAS latency makes no difference is pouring in all the time.

On motherboards that use a dual-channel memory controller, everything about the memory is kept within certain parameters. A year ago you could fool with the memory timings (assuming you're using RAM that can be pushed) and get noticably better performance out of a system that was struggling to get 2 GB/sec memory bandwidth. Now that memory bandwidth is double that, the small margin has no noticable impact on performance -- that's assuming you could get better performance by changing the timings. In all of the testing that I've done so far, the worst performing RAM of the bunch has been Corsair TwinX LL DDR400. The memory bandwidth tests showed that it wasn't significantly better than the standard CL2 XMS or even cheap Apacer DDR333 CL2.5. The 3DMark scores were improved on the LL modules, but the actual in-game benchmarks in Unreal Torunament 2003 (using the HardOCP standardizing utility) showed that there was a very significant decrease in average frame rates with the TwinX -- to the tune of a 40FPS loss. The other memory was getting 151FPS on the Antalus Flyby, but the TwinX got 114. That deficit was consistent for all of the UT2003 tests and the Q3A tests were only slightly better than the standard XMS CL2.

So I tried adjusting the memory timings to their speediest limits on the Asus P4C800, and there was zero change in performance from stock (SPD) settings. That means that -- also contrary to last year's knowledge -- the SPD settings were the best available.

With stock (SPD) settings I found that using a 533FSB processor and an Asus P4C800 motherboard, you could put any functioning, validated PC2100, PC2700 or PC3200 RAM in the machine and it would get the same performance within 100 MB/sec (on a scale of 3400 MB/sec) memory bandwidth, 400 3DMarks (on a scale of 11,400) and the Unreal Tournament 2003 scores were practically identical. A year ago 100 MB/sec was a significant difference, but now it's peanuts.

And just now I got a flyer from Corsair Memory. They claim to be more "stable" than other brands, and show a chart explaining which market each product is intended for. There's also a short article from The RAM Guy saying that the OCZ EL PC3200 couldn't go to the low CAS settings that the TwinX LL could. Of course nowhere in this flyer does it show any performance difference... because in most cases there is none. Not only that, but using DDR400 of any brand on an AMD motherboard can yield *worse* performance than DDR333. As for "stability" I've seen Apacer RAM, which is about a third the price of Corsair TwinX LL, run for hours in overnight burn-in tests with no abnormal outputs, crashes, or other strange behavior. Corsair XMS and TwinX have those aluminum heatspreaders, but I just haven't seen RAM get that hot where it would need one. The larger modules like the double-stacked 1024 and 2048MB sticks do need something to keep them cool because of the close proximity of the chips, but that kind of heat doesnt happen on normal 512MB modules under any conditions that I've seen.

It seems the days of pushing RAM settings to lower timings may be over. Corsair has always made good RAM, but I see no advantage to the average user (or even the "enthusiast" as Corsair says). Overclockers will have to add their own commentary here, as I don't have the equipment to risk for testing.

-Jem
"The proverb says that Providence protects children and idiots. This is really true. I know because I have tested it." -Mark Twain

I wrote this book: www.herotale.com
And here's the next one: www.emeraldblackbird.com
Valour
Black Belt 2nd Degree
Black Belt 2nd Degree
 
Posts: 2893
Joined: Sat Oct 05, 2002 6:24 pm
Location: Rochester, NY USA

Postby atang1 » Wed May 14, 2003 4:58 pm

You experience while true, is difficult to explain, because you didn't have the terminology(words) that can be used to compare your experience.

For instance, you did not have the total memeory used in your system. And you did not have listed your cpu L2 cache size that limits the cached memory size in your system. Then you did not have in comparison, the wait state of your cas setting. The wait state of any cas setting could be changed in your bios. But some bios does not show you the actual turns. The terminology is so many that it is not Ras Cas latency, or dram ns speeds, or wair state of leading edge clocks and trailing edge clocks, and slewrates. Etc., etc., etc. Simple bios lump them all together and set the DDR to an average setting. So, in performance testingh they all have about the same speed. Once you get many motherboards and their bios and see the bios settings in different terminology, you know that your bios did not have the many settings to fine tune your various DDR memories.

Sorry about the overall summation you have, that we can not help you to understand your system without all the terminology and values set in your system..
Last edited by atang1 on Wed May 14, 2003 5:14 pm, edited 1 time in total.
atang1
 

Postby Valour » Wed May 14, 2003 5:09 pm

The test system:

Asus P4C800
Intel P4 2.4B (533FSB 512k L2 cache)
Maxtor 40GB Diamondmax Plus 8 (2MB cache)
Sony 52X CDROM
Vantec Stealth 470 power supply
Windows XP Home (preinstall w/all updates and service packs)

Test software:

SiSoft Sandra 2003 Pro
Unreal Tournament 2003 demo with HardOCP benchmark utility v 1.4
MadOnion 3DMark 2001SE
Quake3 Arena demo (timedemo 1)

If you have suggestions as to how to significantly increase performance by adjusting memory timings, I'll certainly try them.

-Jem
"The proverb says that Providence protects children and idiots. This is really true. I know because I have tested it." -Mark Twain

I wrote this book: www.herotale.com
And here's the next one: www.emeraldblackbird.com
Valour
Black Belt 2nd Degree
Black Belt 2nd Degree
 
Posts: 2893
Joined: Sat Oct 05, 2002 6:24 pm
Location: Rochester, NY USA

Postby atang1 » Wed May 14, 2003 5:58 pm

You have to give the settings of drams in the bios to see what is selected and what is not set by the user. The lack of selectivity is causing all the DDRs useable but not optimized.
atang1
 

Postby Valour » Thu May 15, 2003 2:12 am

atang1 wrote:You have to give the settings of drams in the bios to see what is selected and what is not set by the user. The lack of selectivity is causing all the DDRs useable but not optimized.


Well that's just it -- with new motherboards the SPD chip on the RAM and the BIOS work together to do what you used to have to do manually. I tried a variety of different settings and falied to positively affect performance over the automatic SPD values. In other words if I put CL2 RAM in there, it sets it at what SPD says it is capable of and there is no way to push it manually further without either decreasing performance or failing to POST. Furthermore there seems to be no significant difference in performance between CL2 and CL2.5 when letting the settings be dictated by SPD.

My point is that CAS latency doesn't affect performance to any reasonable degree.

-Jem
"The proverb says that Providence protects children and idiots. This is really true. I know because I have tested it." -Mark Twain

I wrote this book: www.herotale.com
And here's the next one: www.emeraldblackbird.com
Valour
Black Belt 2nd Degree
Black Belt 2nd Degree
 
Posts: 2893
Joined: Sat Oct 05, 2002 6:24 pm
Location: Rochester, NY USA

Postby atang1 » Thu May 15, 2003 9:11 am

One of the secrets of sdram is that 7.5ns drams are installed in PC133 and PC100, the difference is cas 3 or cas 2. They don't change performance. The same dram is also installed on DDR memory.

Your major problem is that even if memory has performance advantage, you have uncached memory(L2 cache limits) which has to prefetch from hard disk. If you see every click has to initiate the hard disk, instead of getting from DDR, you have a very slow system. You have to reduce to the minimum memory size of your operating system to meet the L2 cache limitation to get maximum speed.

We are going to propose to Microsoft to fold the excess dram into multiple banks to cache them all in the cpu L2 chache.
atang1
 

Postby evasive » Sun May 18, 2003 5:02 pm

You are going to propose Microsoft to rewrite a vital part of the OS kernel?
We hate rut, but we fear change.
********************************
System error, strike any user to continue...
evasive
Mobo-fu Master
Mobo-fu Master
 
Posts: 37389
Joined: Sun May 06, 2001 12:01 am
Location: Netherlands

Postby atang1 » Tue May 20, 2003 8:42 am

This proposal is being studied by the Microsoft China operations.

It is absolutely important to bank the total memory to be able to cross point swtich in up to 32 way computers. Hypertransports will be a big improvement in chipsets. Hypertransport is just like dma(5mhz). except is is now faster than FSB, Cpu internal clock, etc.
atang1
 

Postby Aussie » Tue May 20, 2003 11:28 am

Valour here's an article and a review that you might want to check out:-

http://www.theinquirer.net/?article=9532

I know that one is not one of your favourite sites but the article makes some sense.


http://www.lostcircuits.com/memory/ddr400/

This one is a DDR400 review but goes into things quite deeply, in particular I think you will find this page on SPD's enlightening:-

http://www.lostcircuits.com/memory/ddr400/4.shtml



Aussie
Our lives begin to end the day we become silent about things that matter - Martin Luther King
Aussie
Mobo-fu Master
Mobo-fu Master
 
Posts: 10244
Joined: Mon Nov 22, 1999 1:01 am
Location: Mackay, Queensland, Australia

Postby Valour » Tue May 20, 2003 2:04 pm

My Sandra scores between brands, speeds, and CAS latency wasn't terribly different... some of the higher-grade RAM did give consistently higher numbers, but I found the margin of error to be more like 20 MB/sec on my test setup. What was so surprising was the 3D performance -- the in-game benchmarks were practically identical to one another no matter what RAM was used. Only the Corsair LL gave worse scores than the others.

CAS latency just doesn't make a significant difference. The articles above seem to concur:

Bottom line is that CAS latency hardly matters anymore, only in random accesses is there still a benefit, which, however, is at the border of what is, in fact, measurable.


So really what is at heart here for the non-overclocker is this: buy what works and what is recommended by the manufacturer. Now more than ever it is more important to heed that warning. For the first time today I've discovered a motherboard that is not compatible with Corsair XMS (it's the Intel D865GBF), and as I've mentioned before the TwinX LL is not validated for anything other than *TWO* motherboards according to this fact sheet I have from Corsair here.

-Jem
"The proverb says that Providence protects children and idiots. This is really true. I know because I have tested it." -Mark Twain

I wrote this book: www.herotale.com
And here's the next one: www.emeraldblackbird.com
Valour
Black Belt 2nd Degree
Black Belt 2nd Degree
 
Posts: 2893
Joined: Sat Oct 05, 2002 6:24 pm
Location: Rochester, NY USA

Next

Return to System Modification

Who is online

Users browsing this forum: No registered users and 1 guest