Ram timings and what theyre worth

[Marsumane]

Should you sacrifice ram timings to get a higher fsb? (+5 on cpu/mem for +1 on ras and ras to cas timings. NOT cas lat yet) When is it worth it to make your timings worse vs lower fsb for better timings? I assume this is a very touchy subject. Anyone have any benches or examples?

 

[jhurst]

I used to have PC3200 (512X2) in my system that ran at CL2-5-3-2 @ 400mhz (2.8 @ 3.5ghz, 250FSB, 5:4 ratio). I would benchmark using SiSoft Sandra and run the memory bandwidth and get around 5.3GB/s. Now, I have PC4000 (512x2) Kingston HyperX that runs at CL3-8-4-4 @ 500mhz (2.8 @ 3.5ghz, 250FSB, 1:1 ratio), and get benches of around 6.2GB/s. I do not know if the memory bandwidth test measures actual performance, but to me, it was quite a difference.

I would say that timings win out if you are talking about small changes in the FSB. But if you do a serious OC (25%) I would say that MHZ wins out.

 

[Marsumane]

I assumed so also. I mean 5fsb isnt huge even tho it also helps the proc out. It probably doesnt make up for it, but im just thinking that some guru of ocing my have some type of damn good rule for doing such things.

 

[DAPUNISHER]

Tom's ram timing article

Ace's quick comparison of CAS 2 vs CAS 3

PCStats timings vs Bandwidth

 

[DAPUNISHER]

SiSuck means nothing, it'll give great bandwidth scores for running higher 1:1 on P4 but in actual use Members here have proved a divider and tighter timings are faster for most use than 1:1 if the CPU speed can be pushed higher by using the divider to get past the ram's limitation.

 

[jswjimmy]

well on my an35n i see that 400@2.5,3,3,6 is much better that 440@3,2,2,6

 

[DAPUNISHER]

Originally posted by: jswjimmy
well on my an35n i see that 400@2.5,3,3,6 is much better that 440@3,2,2,6

Same CPU speed? Both running synchronous 1:1? If so what benchies have demonstrated this? If it's not the same CPU speed and synch it's just not relevant of course Because those timing differences are minimal and shouldn't be "much better" unless you are having a latency issue caused by asynch mode or running lower CPU speed with higher FSB. Certainly you hit a point of diminshing returns on nF2 but my experience with nF2 does not agree with your's so I'm attempting to determine why. BTW, you will likely find that 8-11 are faster than 6 as a [tRAS] setting link

 

[jhurst]

Originally posted by: DAPUNISHER
Tom's ram timing article

This test doesn't even match up higher MHZ ram (PC4000+) to lower MHZ ram (PC3200-). This just matches timings. Of course the tighter timings are going to own.

Ace's quick comparison of CAS 2 vs CAS 3

This article goes against your argument. It shows that there is virtually no difference between CL3 and CL2. 2-3% at best.....is that worth the extra $150-200 you are going to spend on 2-2-2-5 RAM?

PCStats timings vs Bandwidth

This article shows that PC4000 (3-4-4-8) performs equally with PC3200 (2-2-2-5), losing out by an average of 1% (not 2-3% like the article states, they need to redo their math). It is also a very weak test, testing 2 very limited games (Quake 3, UT2K3 @ 640x480). I would like to see the same tests run on 1) newer games 2) higher screen resolutions.

IMO, paying more for PC3200 2-2-2-5 RAM compared to PC4000 3-4-4-8 RAM is pretty dumb. The articles showed that running standard apps and gaming, results are virtually identical. But if you do need to run your CPU at full utilization, the PC4000 would win out (games are very GPU dependant). Plus, I bet sometime in the future sw manufacturers will smarten up and utilize the extra bandwidth of today's high speed RAM.

 

[Soulkeeper]

I like fast ram timings, but the truth is benchmarks and most real world situations seem to only yeild 5-20% with the majority only taking maybe 6% when comparing the slowest to the fastest timings at the same clock speed. also realize that lowering the latencies will slightly increase memory bandwidth, but not nearly as much as raising the clock speed say another 20mhz might do

 

[jswjimmy]

Originally posted by: DAPUNISHER

Originally posted by: jswjimmy
well on my an35n i see that 400@2.5,3,3,6 is much better that 440@3,2,2,6

Same CPU speed? Both running synchronous 1:1? If so what benchies have demonstrated this? If it's not the same CPU speed and synch it's just not relevant of course Because those timing differences are minimal and shouldn't be "much better" unless you are having a latency issue caused by asynch mode or running lower CPU speed with higher FSB. Certainly you hit a point of diminshing returns on nF2 but my experience with nF2 does not agree with your's so I'm attempting to determine why. BTW, you will likely find that 8-11 are faster than 6 as a [tRAS] setting link

the benchmarks are lower(about 1000points on pcmark 2002) but windows seam to open faster. it might just be me. the ram was 1:1 for both. and realy 8 is faster than 6 thats weaird, ill try it.
Edit: in pc mark 2002
i get 5033 440@3,2,2,8
i get 4353 400@2,3,3,6

so i gess higher mhz is better than lower cas, and i just found out that i can run the ram at 2.5,3,3,6@440mhz

 

[MichaelZ]

tight timings = high frequency (not in sandra but who da hell cares)

 

[edmundoab]

for a small FSB difference like 5 to 10
I think I rather keep my tight timings..

If we're talking like 20Mhz increment in FSB, thats pretty cool to push from CAS 2 to CAS 2.5 I think

 

[jagec]

Originally posted by: DAPUNISHER
BTW, you will likely find that 8-11 are faster than 6 as a [tRAS] setting link

whoa, weird.

/changes tRAS to 11

So the general rule on all the OTHER timings still is, "the lower the better", right?

 

[ShawnD1]

I don't even think those RAM timings do anything. I changed mine from 9-6-6-3 to 5-3-3-2.5 and Sandra said the increase in memory bandwidth (average of 10 tests each case) was only about 2%. The power is all on the frequency it seems.

 

[DAPUNISHER]

Originally posted by: jhurst

Originally posted by: DAPUNISHER
Tom's ram timing article

This test doesn't even match up higher MHZ ram (PC4000+) to lower MHZ ram (PC3200-). This just matches timings. Of course the tighter timings are going to own.

Ace's quick comparison of CAS 2 vs CAS 3

This article goes against your argument. It shows that there is virtually no difference between CL3 and CL2. 2-3% at best.....is that worth the extra $150-200 you are going to spend on 2-2-2-5 RAM?

PCStats timings vs Bandwidth

This article shows that PC4000 (3-4-4-8) performs equally with PC3200 (2-2-2-5), losing out by an average of 1% (not 2-3% like the article states, they need to redo their math). It is also a very weak test, testing 2 very limited games (Quake 3, UT2K3 @ 640x480). I would like to see the same tests run on 1) newer games 2) higher screen resolutions.

IMO, paying more for PC3200 2-2-2-5 RAM compared to PC4000 3-4-4-8 RAM is pretty dumb. The articles showed that running standard apps and gaming, results are virtually identical. But if you do need to run your CPU at full utilization, the PC4000 would win out (games are very GPU dependant). Plus, I bet sometime in the future sw manufacturers will smarten up and utilize the extra bandwidth of today's high speed RAM.

First, I never made an argument Second, you use SiSuck results but attack the testing methodology of established reviewers??? Third, I couldn't agree more with bandwidth being better than timings even on P4 if you can get the CPU speed to be similar and ram speed higher. Fourth, I personally feel timings are more important to P4 than to A64, my experience with A64 is that timings are of little benefit to gaming, and gaming performance is why I bought a A64 instead of P4c.

 

[jhurst]

Originally posted by: DAPUNISHER

Originally posted by: jhurst

Originally posted by: DAPUNISHER
Tom's ram timing article

This test doesn't even match up higher MHZ ram (PC4000+) to lower MHZ ram (PC3200-). This just matches timings. Of course the tighter timings are going to own.

Ace's quick comparison of CAS 2 vs CAS 3

This article goes against your argument. It shows that there is virtually no difference between CL3 and CL2. 2-3% at best.....is that worth the extra $150-200 you are going to spend on 2-2-2-5 RAM?

PCStats timings vs Bandwidth

This article shows that PC4000 (3-4-4-8) performs equally with PC3200 (2-2-2-5), losing out by an average of 1% (not 2-3% like the article states, they need to redo their math). It is also a very weak test, testing 2 very limited games (Quake 3, UT2K3 @ 640x480). I would like to see the same tests run on 1) newer games 2) higher screen resolutions.

IMO, paying more for PC3200 2-2-2-5 RAM compared to PC4000 3-4-4-8 RAM is pretty dumb. The articles showed that running standard apps and gaming, results are virtually identical. But if you do need to run your CPU at full utilization, the PC4000 would win out (games are very GPU dependant). Plus, I bet sometime in the future sw manufacturers will smarten up and utilize the extra bandwidth of today's high speed RAM.

First, I never made an argument Second, you use SiSuck results but attack the testing methodology of established reviewers??? Third, I couldn't agree more with bandwidth being better than timings even on P4 if you can get the CPU speed to be similar and ram speed higher. Fourth, I personally feel timings are more important to P4 than to A64, my experience with A64 is that timings are of little benefit to gaming, and gaming performance is why I bought a A64 instead of P4c.

I wasn't attacking the "established reviewers", i was attacking on how they were being used to back up your argument. It seemed like you were trying to make an argument for timings.....basically what I have concluded is that you aren't going to be able to tell a difference either way, so go either way.......high dollar PC3200, or loose timing PC4000+. I was thinking maybe PC3500 would be good if you could run it at 1:1 @ maybe 245FSB or so?? What can the best Mushkin PC3500 run at with tight timings??

 

[Inappropriate4AT]

Tighter timings beat out FSB hands down, no opinion about it.

This applies, however, to Intel systems only as they can run CPU/MEM asynchrously. There's a really good article that compares an Intel system at around 3.2ghz (3192mhz) 5:4 2-2-2-5(DDR226) vs the same overclock 1:1 2.5-3-3-7(DDR533). I'm pretty sure it's hidden right here on Anandtech somewhere and I've been looking for it for quite some time. I'd appreciate a link from anyone else who knows which article I'm referring to.

SiSoft Sandra is a "synthetic" benchmarking utility, it doesn't actually analyze real world performance, it takes an estimated guess. The best way to truly test performance is through FPS (not 3dMarks) for gamers, and SuperPI/encoding/anti-virus programs for those who do Multimedia. Internet surfers/word processing junkies don't even apply, no offense.

Until someone does extensive testing on AMD systems, I'd imagine a higher FSB would be more beneficial (for DFI/mobile Barton users crunching out 250+ fsb speeds) more or less due to the resulting faster CPU speeds.

It'd be interesting to see the real world results of an AMD system running 2-2-2-5 @ 200x11.5 (2300 mhz) vs more laxed timings @ something like 242x9.5 (2299 mhz). They aren't exactly the same speeds for all those anal people out there, but I'm too lazy to get into more complicated math.

I also agree that the increase in performance is negligible (though much more noticable than most make it out to be) considering the only ram capable of these speeds is no longer in production and will probably cost an arm and a leg to get.