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

[blinky]

yah cas doesnt do much, for 3dmark2001 you need to run it at least 3 times and average the results since the room for error between tests ran at the same settings is so large

[cuddles]

I don't think we need that in depth results, personally. Once you know that going from 2-2-2 to 2.5-3-2 affects performance X amount and going from 2.5-2-2 to 2.5-3-2 affects performance X amount, you can generally know that you've found out how much RAS to CAS will affect performance. I will look into it to see if it affects performance using 4 as well, but if it skews as much as I think it should (if it's the same percentage difference as 2 to 3, then there's no point in extensively benching 3 to 4, since we already know how much it affects it). The point is to really find out how much the timings affect things. I don't see a need to go that in depth with it in order to find those answers. I am however going for RAS to CAS benches tomorrow at some point and hopefully I can get them all done. There will be 6 different sets of timings for investigating RAS to CAS only. Once I find out how CAS, RAS to CAS and RAS affect performance, I will be investigating FSB / Mhz settings and overall clock speeds as well.

Say going from 200 Mhz to 220 Mhz gives you X result, and going from 220 Mhz to 240 Mhz gives you X result as well. At that point we can generally assume going from 240 Mhz to 260 Mhz will give you the same increase, and therefore there isn't really a specific need (in my opinion) for further investigation into it, because we already know what affects we'll see.

With this investigation, I'm after the 'how' of timings, not simply benching and recording the results. Because if you know how certain timings affect things, you can apply that knowledge to any FSB / Mhz setting, rather than consulting a chart for reference.

And thank you all for the encouragement, I'm looking forward to finishing this entire thing and putting a rest to many questions newer people have regarding timings on A64. :thumbsup:

EDIT : You do realize that getting an average for 3 3DMark 2001 runs would take approximately 20 minutes. Multiply that by 6 sets of timings and you have 2 hours of 3DMark 2001 tests alone. I also have to go through all the other benches, which would turn 4 hours of work into 8+. I plan on streamlining the 3DMark 2001 results to LH and LL specifically (to show the differences), if anyone has any other suggestions, I'm open to them. And as I said before, SuperPi 1M will be included with new results to the 100th of a second over an average of 3 runs, and updated for previous results.

[reject]

I ran a few tests. Just the HL2 stress test, everything on the highest details except no AA, or AF @ 1280x960, and 3dmark 01. These results are pretty accurate everything was consistant. They were at 2t because I cannot run 1t stable with only 2.8v and my motherboard does not like my bh5 for some reason either so sorry I couldn't test that. But it's with my sig rig. Also I can't boot at cas 3 because bh5 dosn't like that I've heard and I've never beenable to do it anyways.

HL2 Stress test:

2.5-4-4-7............... 121.5
2-2-2-5................. 128.9

3dmark01:

2.5-4-4-7.............. 18987
2-2-2-5................ 19857

Sorry it is not as comprehensive and thorough as cuddles but it should be pretty accurate results. I'm surprised there was almost a 1000 point gain in 3dmark01 I think this shows very much that timings matter a significant amount. I am going to rerun the test for HL2 at 2-2-2-5 because it jerked a few times for some reason, so I'll rerun that and post my results.

edit: I reran HL2 Stress test 3 times on each setting to make sure they were accurate. I took the highest and the lowest run and here they are.

2.5-4-4-7

highest 122.4
lowest 121.5

2-2-2-5

highest 131.9
lowest 129.3

So there's about an 8-10 fps diff with tighter timings. That dosn't really seem like alot since they are high numbers to begin with but I think that's pretty significant. And keep in mind I was at 1280x960 once you lower the settings and the resolution the gap will most likely increase.

thanks kryptobs2000!
and the Car Low bug is very annoying i have been up at 11 pm saying just one more car low and then going on till 2.30am
i have found that not every app even cares for fsb but all like tight timings, so i make a small sacrifice on those that like fsb and run tighter timings
im gonna do some testing on dividers, tightening timings as i drop dividers. should be interesting casue at 3-4-4 this ram scales up to 280-290 1T but cant run rcd or rp 2 at any speed above 200

[STEvil]

Holst, here are your 754 numbers..

Sisoft/everest only though, and only testing cas and tras with 2x512 OCZ VX.

http://www.xtremesystems.org/forums/...370#post609370


Did those a while back..

[macci]

TCCD wins.

300MHz 2.5-3-3-7 TCCD wins vs 274MHz 2-2-2-5 BH-5 by a couple of fps in system dependant 3D Mark 2001SE tests @ same CPU speed.

I highly doubt that this is the case when both are ran with fastest A64tweaker settings. Looks like every single "page #1" score in the ORB is done w/ BH5 (and that includes Onepagebook's result ).

[fareastgq]

I don't think we need that in depth results, personally. Once you know that going from 2-2-2 to 2.5-3-2 affects performance X amount and going from 2.5-2-2 to 2.5-3-2 affects performance X amount, you can generally know that you've found out how much RAS to CAS will affect performance. I will look into it to see if it affects performance using 4 as well, but if it skews as much as I think it should (if it's the same percentage difference as 2 to 3, then there's no point in extensively benching 3 to 4, since we already know how much it affects it). The point is to really find out how much the timings affect things. I don't see a need to go that in depth with it in order to find those answers. I am however going for RAS to CAS benches tomorrow at some point and hopefully I can get them all done. There will be 6 different sets of timings for investigating RAS to CAS only. Once I find out how CAS, RAS to CAS and RAS affect performance, I will be investigating FSB / Mhz settings and overall clock speeds as well.

Say going from 200 Mhz to 220 Mhz gives you X result, and going from 220 Mhz to 240 Mhz gives you X result as well. At that point we can generally assume going from 240 Mhz to 260 Mhz will give you the same increase, and therefore there isn't really a specific need (in my opinion) for further investigation into it, because we already know what affects we'll see.

With this investigation, I'm after the 'how' of timings, not simply benching and recording the results. Because if you know how certain timings affect things, you can apply that knowledge to any FSB / Mhz setting, rather than consulting a chart for reference.

And thank you all for the encouragement, I'm looking forward to finishing this entire thing and putting a rest to many questions newer people have regarding timings on A64. :thumbsup:

EDIT : You do realize that getting an average for 3 3DMark 2001 runs would take approximately 20 minutes. Multiply that by 6 sets of timings and you have 2 hours of 3DMark 2001 tests alone. I also have to go through all the other benches, which would turn 4 hours of work into 8+. I plan on streamlining the 3DMark 2001 results to LH and LL specifically (to show the differences), if anyone has any other suggestions, I'm open to them. And as I said before, SuperPi 1M will be included with new results to the 100th of a second over an average of 3 runs, and updated for previous results.

I know it takes alot of time to run a test that in depth, but it'd still be nice to see, real data would be better than general guessing... oh well, there's only so much time and we've gotta do other things too like right? hehe.

[cuddles]

I may not be able to get stuff done tonight, due to Christmas and such, I have more to do tonight and tomorrow than I thought. And yes, while I agree it would be nice, as of now I just don't have the time to spend 2 hours watching 3DMark 2001 and re running tests. Maybe at some point, but as of now I can't.

Once we know how changing timings affect things, I think that we'll have a lot more knowledge overall when it comes to asking the question 'Is x-x-x @ xxx Mhz better than y-y-y @ y-y-y Mhz?'.

[bachus_anonym]

TCCD wins.

300MHz 2.5-3-3-7 TCCD wins vs 274MHz 2-2-2-5 BH-5 by a couple of fps in system dependant 3D Mark 2001SE tests @ same CPU speed.

I highly doubt that this is the case when both are ran with fastest A64tweaker settings. Looks like every single "page #1" score in the ORB is done w/ BH5 (and that includes Onepagebook's result ).

also, it's not quite possible to have same CPU speed at 300HTT vs 274HTT. that difference in CPU clocks will also affect benchmark's results.

[Zeus]

I highly doubt that this is the case when both are ran with fastest A64tweaker settings. Looks like every single "page #1" score in the ORB is done w/ BH5 (and that includes Onepagebook's result ).

Not 100% convinced on that one macci.
Didn't OnepageBook do a compare Lobby test with 11x273Mhz 2-2-2-5 vs.10x300MHz 2,5-3-3-7?
Results were about the same, if not slightly TCCD favoured.
I think none of the ORB's #1 page benchers bothered with TCCD RAM and took 2-2-2-x settings to be superior over 2,5-3-3-x for granted, hence the BH-5 (VX?) results one the first page of the ORB.

I have one 512mb stick of TCCD that is 3d proof at 315MHz 2,5-3-3-7 at 3V, i'm pretty sure that can do some damage in 3d mark 2K1.

I could be wrong but i think 2x512mb of 300+MHz TCCD at 2,5-3-3-x shouldn't be overlooked.
I'd love to be proven wrong on that, feel free to do so.

[gclg2000]

This is a good and well deserving thread. I think you should put it all into excel charts and bar graphs. I guess like tomshardware does or something. It makes it easy to get to the nitty and gritty meat of the tests.

This is of course an old old arguement. Many of the TCCD's chips on A64's @ 300 MHZ or so are womping on 250 @2-2-2-5.....there's an equilibrium point and thats' what is required to find.

We all wish and pray for the day that 300MHz in dual channel @ 2-2-2-5 comes...(if ever)

[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.



To answer some questions, I will be using bar graphs closer to the end of everything in order to summarize. As of now I'm concentrating on single timings in order to see their solitary effect on performance. The summary once done will be reasonably large and have bar graphs and all sorts of nice comparisons with FSB / Ram timings

[STEvil]

Doing an awesome job there cuddles, keep up the good work

Do you have UT2K4 or Quake 3?

May be handy to give those two a quick try for a couple benches and see if they react the same as Doom3 and VST are.

[cuddles]

I do have UT2004 installed, Q3A hasn't been installed for ages. I'll try UT2004 at some point, make a 16 bot ONS demo and do a few quick tests to see how it reacts. D3 is more graphical than anything, we all know, and the Source engine is heavy on CPU / Ram it seems, so I used them. They're also the newest type games out, but I will try UT2004 and see how it handles it. If it isn't significant, I'll just mention that it had no effect, if it is significant, I'll go into detailed benching with it

[cuddles]

If anyone could link me to that Process Timer program I've seen when people are running SuperPi to get 100th's of a second times, I'd like to add the results of SuperPi 1M to the previous tests.

[IvanAndreevich]

Good work. I think you should test 10-3-3-3 -> That's what crappy RAM like mine runs.

[dutchman.pt]

Very good work.
It deserves sticky.

[cuddles]

Good work. I think you should test 10-3-3-3 -> That's what crappy RAM like mine runs.

3-3-3 comes along with the RAS Investigation, it will also include combined timings. The question I'll be trying to answer with that final Investigation of timings (next Investigation is using FSB) is :

Say upping tRCD drops you 5%, and upping CAS drops you 2%. Does upping both drop you 7%?

Once this is all done there will be charts with the timings listed to see what performs the best, and the worst. The one suprising thing I've found so far is that 2.5-2-2 > 2-3-2. That really came as a suprise to me

[cuddles]

Why wouldn't it be faster?

tRCD is the timing above all others.

CAS does squat.

Yes, that's clear now. I, like I'm sure many others, previously thought that CAS would have more effect than tRCD, which isn't the case. If I'm telling you things you already know (which I will be for a lot of people) and you have nothing constructive to add, please don't reply. This isn't as clear a lot of people as the colour of the sky is. If you already know what I've said, don't read it, and even better, don't respond.

[Wired-HoH]

I have Corsair PC3200XLPro which has been advertised at 2225 timings. The problem I have, is that most people CAN'T RUN 2225 TIMINGS TO BEGIN WITH. In my case, I have to run at 2336, or I get errors in Memtest and Prime95. When I built my machine, it defaulted to 2225, and it crashed until I figured out what was going on.

Go to Corsair's Forum (www.asktheramguy.com) and apply for an RMA, and don't forget to list all of your specs and settings and how you've troubleshooted so far, or just call up their tech support.

If you notice on the CorsairMicro site, they don't brag about 2225 memory on the main page. It's been replaced with flash memory and upgrades. I wonder why?

Ummm... probably because the XL isn't the newest thing they've done? They came out last May after all. The CMS stuff and that Flash stuff is their newest baby. Come on man, why do you think middle kids are always complaining about a lack of attention?

And I've also run benchmarks at 2225 vs 2336, and there is no differance.

How'd u run benchmarks if you were crashing out at it, or did I read that wrong?

[STEvil]

Here you go cuddles, found process timer for you.. took me a while to dig it up..

http://xtremesystems.org/forums/show...&postcount=777

[cuddles]

Ahh, thank you. I searched Google for about 20 minutes before I gave up the search. Will update with SuperPi results on Friday morning (4AM or so EST) and hopefully also the quick RAS investigation, as I don't expect to see much of a performance difference.

[macci]

Didn't OnepageBook do a compare Lobby test with 11x273Mhz 2-2-2-5 vs.10x300MHz 2,5-3-3-7?
Results were about the same, if not slightly TCCD favoured.

Yes he did but he forgot to 'tune' the BH5 to max settings w/ A64 tweaker (those tweaks dont lower the max OC clockspeed at all).
'untuned' Bh5 at 273 is like 'maxtuned' bh5 at 253 => check this thread for more info
If TCCD would actually be faster you'd see it on top of the ORB

[macci]

It was clearly visible in his screenshot.

[jmke]

This thread actually deserves to be made into an article and published.
Review sites are probably scared to do it, because they don't want to make anyone mad.

why would they be scared?

I have Corsair PC3200XLPro which has been advertised at 2225 timings. The problem I have, is that most people CAN'T RUN 2225 TIMINGS TO BEGIN WITH.

faulty ram, RMA it.

[reject]

cool, cant wait till its done
i wanna know whether to ditch this 3-4-4 or use the money to upgrade something else