How memory timings on A64's affect performance - The truth.

[cuddles]

December 28, 2004 : Added RAS to CAS Investigation. Updated CAS results with LL / LH scores and removed CPU Arithmetic. SuperPi to be added shortly, was unable to find a Process Timing program to measure SuperPi calculation time within 100th's of a second. Updated first post with benchmark methods and details, as well as an A64 Tweaker screenshot.


Ever hear this : "Ram on Intel loves Mhz, ram on A64's love tight timings"? I know I have, and I had never, ever seen any solid evidence to state this as fact. Now, when you're dealing with something like memory timings, a lot of things come into play. Mobo, BIOS, ram type, and of course the CPU. There are also many facets of performance. You of course have your synthetic benchmarks, your benchmarks from the competitive realm (AQ3, 3DMark) as well as your pure real world gaming performance. I've seen on a lot of forums people asking about how ram timings will affect their performance, whether it be in competitive benchmarks or real world gaming. Not finding any hard evidence, I decided to look into this topic.

By no means do I consider myself an expert at all in the Ram department. At the time of me typing this, I've only been actually overclocking my computer for about 1.5 months. I've been reading about overclocking for maybe 4 months. In the following sections, I'll take a look at CAS, RAS to CAS and RAS effects on performance, as well as a compilation at the end including FSB and performance ratios and other data to help you sort out where you are most likely to perform best (220 3-2-2 > 200 2-2-2?).

The test setup is as follows :

S939 AMD FX-53 (200x12, stock)
2x512MB Crucial Ballistix DDR400 (running in Dual Channel)
MSI K8N Neo2 Platinum (1.3b6 BIOS)
x800 Pro flashed to XT PE @ 570 / 590
2x74GB Raptors in Raid 0
Windows XP SP2
Audigy 2 ZS Platinum Pro

Notes about the conditions of the benchmarks :
- I tried to keep all tests on even playing ground. All of the same services were running in the background for all tests at all settings.
- If a benchmark looked off, or fishy, I ran it more than once to make sure my results were exact. There will always be a slight margin of error, which can be attributed for some differences.
- VSync was never enabled.
- The ATI Control Panel was set to High Quality for all tests.
- For the Competitive Benchmarks I simply ran the tests from beginning to end. No tweaks, no tricks, no nothing.
- In 3DMark 2001 I only use Lobby High and Lobby Low to show results. Originally I ran through the entire benchmark, but margin of error in 2001 is very large. At the advice of the 2001 Pro's, LL and LH were deemed as the most RAM / CPU dependant tests, and therefore those are the only scores you will see.
- A64 Tweaker shot for specific settings on all benchmarks can be found here.

I'm breaking this down into 3 subsections, Synthetic Benchmarks, Competitive Benchmarks and Real World Performance.

Here are the settings for the gaming benchmarks :

Counter-Strike : Source Visual Stress Test (Ran once, running twice gave identical results)
Low :
640x480
All settings on Low
Bilinear Filtering, No Reflect

High :
1600x1200
All settings on High
4xAA / 16xAF / Reflect All

Doom 3 - timedemo demo1 (Ran twice, took second score)
Low :
640x480
Lowest Quality
All Advanced Settings disabled

High :
1600x1200
Ultra Quality
All Advanced settings enabled (except for AA and VSync)

And without futher delay, here are the results!

EDIT - Quick Links
Investigating CAS
Investigating RAS to CAS (tRCD)

[cuddles]

Investigating CAS


Synthetic Benchmarks

200x12 / 2-2-2-10 :
Everest Latency Benchmark : 47.2ns
Sandra Soft 2005 Bandwidth : 6070 / 6027

200x12 / 2.5-2-2-10 :
Everest Latency Benchmark : 50.8ns
Sandra Soft 2005 Bandwidth : 6055 / 6012

200x12 / 3-2-2-10 :
Everest Latency Benchmark : 53.1ns
Sandra Soft 2005 Bandwidth : 6048 / 5994

As you can see, going from CAS2 to CAS2.5 will lose you about 3ns in latency (7.6%). Going from CAS2.5 to CAS3 will hit you for about another 3ns.

And how does this affect your bandwidth? Not much overall. The total difference from CAS2 to CAS3 is only a 22 / 33 decrease (0.25% / 0.25%). And from CAS2.5 to CAS3 it's only 1%! That definately isn't much in the grand scheme of things, as you'll see when we get to real world gaming.




Competitive Benchmarks

200x12 / 2-2-2-10 :
3DMark 2001 : 351.3 / 156.2
3DMark 2005 : 6312
Aquamark 3 : 77661

200x12 / 2.5-2-2-10 :
3DMark 2001 : 344.5 / 153.3
3DMark 2005 : 6304
Aquamark 3 : 76749

200x12 / 3-2-2-10 :
3DMark 2001 : 338.2 / 150.9
3DMark 2005 : 6309
Aquamark 3 : 75755

(Note : Most competitive benchers on the A64 platform make use of A64 Tweaker by CodeRed, I'd just like to note that I did not use that utility at all for any of these benchmarks. CAS timings were set in the BIOS. Lobby Low and Lobby High are the scores shown. A64 Tweaker shot for specific settings on all benchmarks can be found here.)

Interesting results we see here. Competitive Benchmarks really do get affected by things as specific as your ram timings, which most competitive benchmarkers already know I didn't use 2003 for this test because it's essentially in between 2001 and 2005 as far as CPU / Ram dependance.

3DMark 2001 showed a large decrease when switching CAS timings. Going from CAS2 to CAS2.5 decreased Lobby Low by 6.8 (1.97%)FPS ! As benchers know, this is a huge decrease. Going from CAS2.5 to CAS3 dropped it another 6.3 FPS (1.86%). These may seem like nothing to most people, but when it comes to competitive benchmarking, almost 2% is a huge, huge difference. In Lobby High we saw the same drop happen, we lost 5.7 FPS (3.51%) going from CAS2 to CAS3.

3DMark 2005 showed slight differences which are within the margin of error, so we can basically ignore all those tests. This just goes to show you that 3DMark 2005 is essentially entirely GPU as of now. In 4 years it will probably just as CPU bound as 3DMark 2001 has become.

Aquamark 3 showed interesting results as well. I never thought Aquamark 3 was this dependant on things as specific as ram timings. We see that from CAS2 to CAS2.5 we drop an astounding 912 (1.18%) points. Further, from CAS2.5 to CAS3 we drop another 994 (1.31%) points. The total difference from CAS2 to CAS3 is 1906 (2.52%) points. That is a huge difference, if you ask me when it comes to competitive benchmarking.

And now, we move on to the true test, real world performance.



Real World Performance

200x12 / 2-2-2-10 :
VST Low : 198 FPS
Doom 3 Low : 144.2 FPS

VST High : 102.74 FPS
Doom 3 High : 61 FPS


200x12 / 2.5-2-2-10 :
VST Low : 198 FPS
Doom 3 Low : 143.7 FPS

VST High : 101.7 FPS
Doom 3 High : 61 FPS


200x12 / 3-2-2-10 :
VST Low : 193 FPS
Doom 3 Low : 142.2 FPS

VST High : 101.07 FPS
Doom 3 High : 61 FPS


Very interesting results we see here. Even when we're running at lowest settings, taxing the CPU / Ram much more so than the video card, we see almost the exact same numbers throughout.

Doom 3 Low barely budged. It isn't within margin of error, it is indeed because of the CAS timings, but a total of 2 FPS difference from CAS2 to CAS3 is pretty significant when it comes to showing that timings don't matter that much in real world gaming.

Doom 3 High was entirely GPU. Frames don't change at all, why? Well, something as insignificant as ram timings won't change anything if all the work is being put on your GPU, so we see absolutely no performance decrease.

VST Low gave almost the same results as Doom 3 Low. Now, on CAS3 it did seem to drop more than from CAS2 to CAS2.5, and I did rerun the test to make sure, and it hit 193 each time. Even so, a 5 FPS drop when you're dealing with settings that will never be used real world don't have much weight, in my opinion.

VST High also gave interesting results. This is real world, this is what you'll see while playing. From CAS2 to CAS2.5, we lost 1 FPS. Oh no!! Not a whole frame! Going from CAS2.5 to CAS3 dropped us not even a single frame per second, 0.63 if you want to get technical. I think we can all agree that that means nothing in terms of real world performance. Overall, going from CAS2 to CAS3 dropped us approximately 1.63 FPS.



Conclusions

Well, overall, I think we got some interesting results. On forums a lot I'll hear people ask 'Should I get this 2.5-3-3 ram or this 3-3-3 ram?'. In most cases, the 2.5-3-3 ram will be a decent amount more expensive than the 3-3-3 ram. And most people get the answer of 'Go for the 2.5 ram, it's worth it'. I think after the tests given above, we should really start questioning that.

For Competitive Benchers
I don't think I need to say much here. If you bench for sport, you already know what I've told you. Timings do matter, and they matter a reasonable amount. The difference at the same Mhz between CAS2 and CAS2.5 even is a reasonable amount. The difference between CAS2 to CAS3 is even more pronounced, and as we all know, 500 points in any benchmark can make or break a world record. Paying the extra $50-$100 for CAS2 ram instead of CAS2.5 is definately worth it. But as I said, if you're a bencher, you already know all of this and probably have some nice sticks of BH5 or TCCD at your side already, so I don't think I need to say much more on this front.

For Gamers (especially those on a budget)
Here's what it really all comes down to. If you're a gamer with a newer PC, you're most likely on the A64 platform. The question of timings comes into play a lot when deciding your PC components. I even previously would tell people to spring for the CAS2 over the CAS3, thinking it would indeed help with performance. Keep in mind, this is specifically dealing with CAS only. If you have the choice between 2-2-2 and 3-2-2 ram (which won't happen often), then I think you can see that the extra money usually isn't worth it if you're an average gamer. After the tests above, I really have to question myself. Is it really worth it? For a budget gamer, $50-$100 is a pretty decent amount of cash to smack down for some CAS2 ram as opposed to some CAS3 ram. Is it truely worth it? I guess that's up to the gamer themselves, but overall I think we can come to the conclusion that the answer is no. $50 extra for 1 FPS under normal gaming conditions is not something most people would spring for. Even when it's entirely CPU / Ram dependant, we barely see a 3 FPS difference. This is entirely negligable, I think we can agree. I myself was surprised at how little difference in the real world ram timings had an effect. Update - Keep in mind, this is CAS only.



Final Thoughts

Well, when it comes to CAS only, I think it's clear that for real world performance, CAS doesn't mean a thing. For benching, of course it does. And for synthetic benchmarks, you will see a performance decrease, but it's extremely minimal (well under 100mb/s from CAS2 to CAS3). Scroll down for more benchmark results regarding other timings. But as a final answer for CAS only, it has no real world performance decrease when changing from CAS2 to CAS3 (keeping all other timings equal).

[cuddles]

Investigating RAS to CAS (tRCD)



Synthetic Benchmarks

200x12 / 2-2-2-10 :
Everest Latency Benchmark : 47.2ns
Sandra Soft 2005 Bandwidth : 6070 / 6027

200x12 / 2-3-2-10 :
Everest Latency Benchmark : 50.9ns
Sandra Soft 2005 Bandwidth : 6052 / 6006

200x12 / 2-4-2-10 :
Everest Latency Benchmark : 52.9ns
Sandra Soft 2005 Bandwidth : 6027 / 5983

When it came to CAS only, we saw very minimal performance loss (22 / 33 decrease). When it comes to tRCD2 to 4, we see a 43 / 44 (0.71% / 0.74%) decrease overall. Going from 2 to 3, we only see a drop of 18 / 21 total. When it comes to memory bandwidth, this isn't a huge drop, but it is noticable when it comes to synthetic benchmarking.

The latency also suffers a reasonable amount, moreso than with CAS only being the variable. We drop 4ns from 2 to 3 and another 2ns from 3 to 4 for a total drop of 6ns. That's a reasonable amount when it comes to your ram timings. Much more pronounced than we saw with CAS changing.

With CAS from 2 to 3, we had a total drop of about 0.25%. With tRCD from 2 to 4 we see a ~0.72% decrease. More than twice the performance difference in comparison between the two settings when it comes to bandwidth.


Competitive Benchmarks

200x12 / 2-2-2-10 :
3DMark 2001 : 351.3 / 156.2
3DMark 2005 : 6312
Aquamark 3 : 77661


200x12 / 2-3-2-10 :
3DMark 2001 : 340 / 151.1
3DMark 2005 : 6316
Aquamark 3 : 76573

200x12 / 2-4-2-10 :
3DMark 2001 : 332.3 / 147.7
3DMark 2005 : 6305
Aquamark 3 : 75668

Looks like we have some more 'splainin to do.

3DMark 2001 showed almost the same decrease as before. With CAS only we had a decrease of almost 4%. Here, going from tRCD 2 to 4 dropped us 19 FPS (5.7%) in LL. In LH with CAS we saw a total 3.51% drop. With tRCD from 2 to 4 we see a drop of 8.5 FPS (5.75%). The differences between CAS and tRCD so far show us that tRCD effects performance close to twice as much overall in 3DMark 2001.

3DMark 2005 is almost not worth mentioning. This should probably be dropped from future benches, but I left it in just to show that Ram timings mean next to nothing in 2005. Going from CAS2 to 3 will drop maybe 2 points off your score, and going from tRCD 2 to 4 will drop you about the same.

Aquamark 3 also showed us more of a decrease than with CAS. While we were seeing a total of 1906 (2.52%) points, going from tRCD 2 to 4 drops us about 1993 (2.63%) points. While less drasatic than the CAS2 to 3 drops we saw previously, tRCD only increase your performance decrease (yeah, that makes sense :p ) by about 0.1%. Not entirely significant, but the drop is there.

Overall we can see that 3DMark 2001, as usually is very picky about ram and doesn't like the tRCD being upped as much as the CAS. With AQ3, we see that it doesn't really have a preference as to what is being changed. Upping CAS or tRCD will get you about the same decrease.


Real World Performance

200x12 / 2-2-2-10 :
VST Low : 198 FPS
Doom 3 Low : 144.2 FPS

VST High : 102.74 FPS
Doom 3 High : 61 FPS

200x12 / 2-3-2-10 :
VST Low : 195.84 FPS
Doom 3 Low : 142.8 FPS

VST High : 102.27 FPS
Doom 3 High : 61 FPS

200x12 / 2-4-2-10 :
VST Low : 188.37 FPS
Doom 3 Low : 140.1 FPS

VST High : 102.65 FPS
Doom 3 High : 61 FPS

Doom 3 Low really isn't worth mentioning. Once again, we about a 2 FPS difference between timing changes. Woopdee.

Doom 3 High is what you'd notice playing. And yes, you wouldn't notice a thing. Rock hard 61 FPS all throughout, it just doesn't budge.

VST Low was also worthless. There was a decrease of about 5 FPS total from tRCD 2 to 4. Not huge at all, considering no one would even play at these resolutions. It does show that the ram timings are affecting gameplay under these specific circumstances, so let's see how it will really affect things.

VST High shows us... a suprise!. With CAS2 to 3 we saw a 1 FPS decrease. With tRCD 2 to 4 we see nothing change at all. They're all within margin of error, so we get the exact same results with tRCD 2 to 4 throughout. Not something I was expecting, but I'm not particularly suprised either. CS: Source just doesn't care about tRCD at all.


Conclusions

I think we saw what we expected. There is a performance decrease with synthetic and competitive benchmarks when increasing RAS to CAS (tRCD). In the competitive benchmarks, it is very significant, and it also is in the synthetic benchmarks, especially in latency. But with real world, we see once again it makes essentially no difference. Remember, this is tRCD only!

For Competitive Benchers

What we see here is that you will see a bigger decrease when changing tRCD compared to CAS. Going from 2 to 3 on tRCD hits you for almost 1000 AQ3 marks and drops your LL by ~10 and LH by ~5. If you have a choice between 2.5-2-2 and 2-3-2, I think it's clear that you'll want 2.5-2-2. This actually suprises me, but when looking at FPS comparisons we see that upping tRCD one step hits you for about twice the performance that upping your CAS does. I personally find this suprising, as I would have naturally thought 2-3-2 would outperform 2.5-2-2, but as we see here, that isn't the case.


For Gamers

Well, here's where most of the populace is, even the hardcore OCers are usually primarily gamers going for the most out of their system and their budget. Once again in Doom 3 we see no performance drop when playing with tRCD, and we also see nothing change in the VST at all. In fact, tRCD affects CS: Source less than changing CAS does. And when I say 'affects', we're talking about 1-2 FPS, which is practically nothing. $50 for 2-2-2 over 2-3-2 clearly isn't worth it. Will it be worth it with 2-2-2 vs 2.5-3-2 or 2.5-3-3? We'll find out shortly

Final Thoughts

Another timing down, another fable perhaps destroyed. When it comes to real gaming, tRCD matters just as little as CAS does. Keep in mind, these are singular variables, with only one timing being changed at a time. The entire story could change with CAS as well as tRCD changing, but that's something we'll have to wait to see.

The bottom line with RAS to CAS (tRCD) : In real world performance, going from tRCD2 to tRCD4 has no performance decrease.

[cuddles]

Saved for future results - Investigating RAS.

[cuddles]

Saved for future results - Mhz differences and what Mhz increase is required to make up for timing differences (i.e 220 2.5-3-3 = 200 2-2-2?)

[viccyran]

Nice post.

Now.. I know you worked hard enough, but we should probably compare mhz! Try to see the best buy for all those gamers out there

Good work

[cuddles]

I will be definately looking into if the results are different at higher Mhz. I will be updating the posts as I get more results, don't you worry

[Soulburner]

Nice testing, however my only gripe are the actual timings you used.

When people talk about memory timings they usually talk about the differences between 2-2, 3-3, and 4-4 timings. You tested by only changing the CL rating which has a very minimal impact on performance.

Can you test with different timings? (ex 2.5-3-3-7, 3-4-4-8, this is where you will see how much timings matter)

[cuddles]

Nice testing, however my only gripe are the actual timings you used.

When people talk about memory timings they usually talk about the differences between 2-2, 3-3, and 4-4 timings. You tested by only changing the CL rating which has a very minimal impact on performance.

Can you test with different timings? (ex 2.5-3-3-7, 3-4-4-8, this is where you will see how much timings matter)

You may be asking, 'Well what about the other timings? What effect does 2-2-2 vs. 2-3-2 have? What about 2-2-2 vs. 2-3-3 or 2-2-3?'. I'm asking myself the same question, and I do plan to go further into it to find out how other timings affect A64 systems, if at all. Specifically getting down to 2-2-2 vs 2-3-2 and 2-2-3, to really see if there's any effect at all in real world gaming or benchmarking. I also want to look into different Mhz settings and see if it really makes a difference when combined with different timings. I will update the thread as I do so.

All in time This entire thing (as well as write up) took about 4 hours, I will get to specific tests very shortly and update the posts with the new information.

[Dojo | Warlord]

Mabye this thread deserves a sticky?


GJ

[Mtemel]

nice work man

For me its been impossible to find memory that will actually go above 258 fsb on 2.5,3,3,7, ive had 3 lots of mem and none go above 258

[reject]

f***n awesome that must have taken ages

when i saw u did cas first i thought it would make no diff but 2k1 really likes low cas!
very good to know since its most peoples fav bench
and the games was what i expected, still good to know i dont lose performnce gaming at my 24/7 settings

[STEvil]

lol.. i'll bet it did.

Took me quite some time to do 1.5/2/2.5/3/2-2-0-/5/10/13 numbers for just sisoft and everest...

[MaxxxRacer]

i knew that 2001 was cpu dependent, but omg... now i know why i get garbage scores... cant wait for this rma on my winnie and my ocz ram. should be able to break 30k.

cuddles, that is a great article. I say sticky. But i would sticky it in the amd forums as this is an amd only post..

[reject]

the amd forum is chockers, the memory timings forum would be better

[Kunaak]

personally, I never found any ram that worked the same way, the ram I had before it did.

all ram seems to be very different.
while my kingston 3500 can do up to about 280 at cas 2-2-2-5, you will see I always run it at cas 2-2-2-7, or sometimes 6.
the tras on my Kingston never seems to do anything better at tras 5 or less, but at tras 7, my bandwidth seems to be at it's magic spot, just about 20 mbs or so higher then when using tras 5, and always gets me alittle more points, when used on A64's.

most people would tell me to go for cas 2-2-2-5 when benchmarking, but with this ram, and having 2 years of using this exact ram, I can tell you, the magic spot for my ram is tras 7.

however my geil ultra X, seems like great stuff, does about 260 at cas 2-2-2-5 with just a hair over 2.9 volts easy.
great little ram, but it doesn't seem like cas 2 much, and even gives me a nice half second boost in pifast, if I ran at 2.5-2-2-5.
even better if I ran 2.5-2-2-0.

tras is almost useless for my geil TCCD stuff, and cas latency seems to only hold it back from the higher scores in 3dmark and pifast, when past about 255, and it doesn't respond to voltages at all past about 3 volts.
very weird ram, but hard to complain when you have ram that will run cas 2.5-2-2-5 at 260 with only 2.9+ volts all day, without any issues.
it's not gonna replace my BH5 kingston for benchmarks, but it does have a spot in my hardware pile, waiting for some new hardware.

I think when it comes to ram, the best way to know where you ram the best, isn't with a guide, it's with a few hours of testing your own ram.
sounds like a pain, but once you know, you never have to guess what you ram can and cannot do again...

atleast, till you get a new motherboard

[Wingless]

// edited by Jupiler for flaming.

[MrQ3W]

What is the highest level of education you've attained Wingless?
Hope you get what you so hope for..

[DoGMaN]

Very nice information cuddles, good work. Though I agree with SoulBurner. I think it would have been more helpful if you would have done these tests with different tRCD rather than CAS. I think you would find that it makes a little more difference on the A64 platform.

[Duka]

Very well done!!!
However as it has been said previously by Soulburner, its not alone CAS 2, 2,5 or 3 but also the other timings when the "gamers" are considering wheather they should buy cheap value ram or expensive ram.

For instance a comparison between 2.2.2.10 versus 2,5.3.3.7 and perhaps even worse timings.

Edit: Was a little to fast on the bit about grammer...

[Dunk]

He really likes that nana

Nice review but as others said it would have been more interesting to focus on different timings. Here's my best results for tweaked timings with my PQI 2x512MB DC kit at stock for my winchester 3500+ on my neo2 plat:

Column Address Strobe # (CAS) = 2.0
Row Cycle Time (Trc) = 7
Row Refresh Cycle Time (Trfc) = 9
RAS# to CAS# delay (Trcd) = 2
Row to Row delay (Trrd) = 2
Min RAS Active Time (Tras) = 10
Row Precharge Time (Trp) = 2
Write Recovery Time (Twr) = 2
Write to Read Delay (Twtr) = 1
Read to Write Delay (Trtw) = 1

spent a long time messing with A64 tweaker before I settled on those and did a lot of testing

[kryptobs2000]

I bet you will definatly see more of a performance difference if you did with soulburner suggested I agree as well. Also, I bet it'll make alot more difference when running at highermhz say 250mhz 2-2-2-5 vs 250mhz 3-2-2-5 since it will have more bandwidth of course. Can't wait till you finish ur testing I've done a few benchmarks but nothing extensive.

To wingless, high fsb does make a pretty noticable difference, I don't see how you can conclude anything related to fsb based on these tests. If you don't like it here your more than welcome to leave.

Great job cuddles

[HermS]

Wow thanks wingless you've changed my whole perception of overclocking... you're a genius

Interesting experiment man, well done!

[mcnbns]

Yes. Ban wingless if only for his risking inducing epileptic seizures to all who read this wonderful thread.

Awesome info, cuddles, and thanks for your hard work.

[macci]

Good idea to show the realworld difference but like already mentioned the timings that actually matter are Trcd and Trp (w/ A64 there is great number of other timings too). CAS lantency hasn't been important in a while.

While doing the mem latency comparisons it might be a good idea to show how much CPU clock speed makes a difference in gaming performance. Example: is it better to have 2200MHz w/ 3-4-4-8 timings or 2000MHz w/ 2-2-2-5 timings.