Timings: tRAS 11 is superior to 5

[AsiLuc]

For a while now I've been contemplating this, and I'd like your opinion on this one. It's about this 5 that's been nagging me. Maybe you're all past it, if so, tell me. If not, this'll get interesting..

You guys like performance, right? When we talk memory, it's usually like "I run 2-2-2-5!", "Can't run 2-2-2-5? Get some BH-5"! etc....

This seems to imply that 2-2-2-5 yields better performance than, say 2-2-2-11. But is that really so?

First, start reading this (http://www.mushkin.com/mushkin/pop-up/latencies.htm). A little bit about latencies written by Mushkin.

I assume you've read it all.

Now, check this (http://www.madshrimps.be/?action=getarticle&articID=133). An intelligent article by MadShrimps.

If you're not into reading (http://www.madshrimps.be/articles/AMDmemorytiming-piotke-2237.gif).

Let me summarize: 2-2-2-11 beats 2-2-2-5. Why? read the Mushkin story.

So what do you think?

[TheDogFather]

Thinking hurts but look at my sig.

TDF.

[AsiLuc]

Yo TDF, I understand you're running 2-2-2-11 and I like it, but what do the 13 and 15 mean? How should I 'read' your 11-2-2-2-13-15?
Does everybody know this?

[TheDogFather]

DFI have these extra settings, to tell the truth I dont renember what they are. The extra 2 is the last 2 split into 2 then theres the 13 and 15 whatever they are. :D

The 11 thing is common knowledge though.

TDF.

[AsiLuc]

Thanks for the info TDF. I've been pretty funny then :P
At least now I know it's known, but still not totally.
Oh and btw, thinking doesn't really hurt.

[Orak]

many use 10 insted of 11 on NF2 boards, dont know why.. didnt get time to check it out when i had NF2 .. ;)

[Tony]

TRAS is the time you set that allows a page to be read.if the time is to short the page will be partially read or skipped and the controller will then back peddle to read the page again.If you tune TRAS you will get better performance..so no re reads and the pages are only opened once.

Hope this helps.

[AsiLuc]

Yep you're right bigtoe.

[Demon_Hunter]

so that is true, 2-2-2-11 is better than 2-2-2-5/6 ???

in NF2 boards is better 2-2-2-10 !?!?!?!?!?

see ya

[jmke]

2-2-2-11 is better on nForce2 boards ONLY :)
some ram might like a Tras of 12 or 10 better, you'll have to check to find it out

OCZ's EB likes -8 best I we believe Anandtech's story

[XanderF]

It IS nForce2-specific...and it DOES depend on the board/chipset.

I did a bit of tweaking on mine going up and below that setting. IIRC, 8 and 10 were identical in performance, so I went with 10 since it would allow a higher memory clock.

[Sir Lancelot]

Originally posted by AsiLuc
For a while now I've been contemplating this, and I'd like your opinion on this one. It's about this 5 that's been nagging me. Maybe you're all past it, if so, tell me. If not, this'll get interesting..

You guys like performance, right? When we talk memory, it's usually like "I run 2-2-2-5!", "Can't run 2-2-2-5? Get some BH-5"! etc....

This seems to imply that 2-2-2-5 yields better performance than, say 2-2-2-11. But is that really so?

First, start reading this (http://www.mushkin.com/mushkin/pop-up/latencies.htm). A little bit about latencies written by Mushkin.

I assume you've read it all.

Now, check this (http://www.madshrimps.be/?action=getarticle&articID=133). An intelligent article by MadShrimps.

If you're not into reading (http://www.madshrimps.be/articles/AMDmemorytiming-piotke-2237.gif).

Let me summarize: 2-2-2-11 beats 2-2-2-5. Why? read the Mushkin story.

So what do you think?


Does that apply also to Intel systems running at 5:1 ??

[Nevets]

I've been OC my mem and have gotten some unexpected results. I have found a "sweet spot" in my system at 11X232 (2559) (auto) running 2-2-2-5. When I bchmk using Prime95 I am getting a memory best time at 21.6 ms whereas everyplace else I go, I cant get any faster than 26.2 ms.

Now I am assuming (hate that word) that this is a good place to be for memory and FSB because I got a 3D2001SE score of 12694. (Got pictures but I can't repeat it).

Is this a "norm" for memory or is there another explanation for this?

I did do a Vdimm mod and am at 3.3V, Vcore @ 1.825V (only because the MOBO has loose voltage control) and 1.8 on the NBridge.

If it is a sweet spot, sthe next step will be to repeat the 3D and torture test. I could loosten the timings to go for 2-2-2-11 and see what happens. Any other ideas?

[BuickBeast]

11 works best for my nforce2 abit.

[krampak]

-11 was the best value for my NF7S but in the Infinity it was some faster 8 :) With Geil 433 Ultra series. Although I tried -3 and was aprox. as fast as 11 in nf7s.

[STEvil]

2-2-2-2-11-9-12-e-e-f-e-1-3-3-3-2-1-3-auto-auto-auto-auto-auto-auto

These infinities have so many timings its redicuous! :D

[iddqd]

TRAS is the time you set that allows a page to be read.if the time is to short the page will be partially read or skipped and the controller will then back peddle to read the page again.If you tune TRAS you will get better performance..so no re reads and the pages are only opened once.

Hope this helps.
Then why is the "optimal" value cas-dependant? For example - nF3. tCl=2, optimal tRas=10. tCl=1.5 optimal tRas=0.

[enzoR]

are you sure tRas 0 is optimal for cl1.5?

[KaEL]

Recently, I've been testing(a lot) BH-5 with my FX-55,

For Athlon64 FX-55 memory controller TRAS = 8 clocks delivers better performance on real life.

2-2-2-8 > 2-2-2-5, 2-2-2-6, 2-2-2-11.

regards,

KaEL

[Lufusol]

Now imagine someone closes the book you are reading from in the middle of a sentence. Right in your face! And does it over and again. This is what happens if tRAS is set too short. So here is the really simple calculation: The second burst of four has at least to be initiated and prefetched into the output buffers (like you get a glimpse at the headline in a book) before you can close the page without losing all information. That means that the minimum tRAS would be tRCD+CAS latency + 2 cycles (to output the first burst of four and make way for the second burst in the output buffers).
So if you're lucky enough to have chips that run 2-2-2, obviously 5 is not ideal, but by that formula, 6 (as a minimum) is. Can someone explain the theory behind pushing all the way to 11+? Or has this number been found ideal through trial and error? I assume every chipset's responsiveness to one value or another is different...


...there are very few sticks that can do CL2 2-2-5 timings. Most ram even won't do "Ras To Cas Delay : 2" and "Ras Precharge : 2". Mostly it's the Ras to Cas that's holding you back, set it to 3, or higher, and problem solved.
Then why did he test timings of 2-2-4-11 and not 2-4-2-11 in the benchmark? If someone can explain this as well, I'll give you a gold star. I sure have no idea.

-Luf

[XanderF]

So if you're lucky enough to have chips that run 2-2-2, obviously 5 is not ideal, but by that formula, 6 (as a minimum) is. Can someone explain the theory behind pushing all the way to 11+? Or has this number been found ideal through trial and error? I assume every chipset's responsiveness to one value or another is different...

Trial and error is the ticket.

As I noted above, timings of 11+ (I'm using 13 on the motherboard I have at the moment) seem to be ideal for nForce2 boards *only*.

I read on a forum somebody's guess that this is due to the L3-cache-like-thing that the nForce2 chips have. If you recall, there was some talk about this "Dynamic Adaptive Speculative Pre-processor" (http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=1654&p=4) at the nForce2 launch, without any specifics. Anyway, like I said, the running theory at the moment is that the very high tRas is needed for the DASP functionality to really shine.

[sjohnson]

ymmv, one application may run better at one timing than another, your board might "like" one timing better than another.

Good sources of information, all those links, but what works best for your needs on your board is what counts.

What I mean is, don't discount testing at other than "the best" advertised settings just because you respect those who advertise :)

[Aguanteboca]

A little discussion about this very same topic...
http://www.dfi-street.com/forum/showthread.php?t=9114

[Ubermann]

Use the value that gives best performance and is stable without error.
And ignore that report.

[Mardok]

quick test:

http://members.lycos.co.uk/m4rd0k/xs/2-2-2-11.png
http://members.lycos.co.uk/m4rd0k/xs/2-2-2-8.png
http://members.lycos.co.uk/m4rd0k/xs/2-2-2-5.png

[psy^]

I find 2.5-4-4-0 1T yields best performance for my 3200XL sticks @ 260Mhz :)

Tras of 12 is maybe 10MB/sec slower in synthetic benchmarks?
Not that i can tell the difference, but it just looks cool running 0 :p:


TRAS is the time you set that allows a page to be read.if the time is to short the page will be partially read or skipped and the controller will then back peddle to read the page again.If you tune TRAS you will get better performance..so no re reads and the pages are only opened once.

Wouldn't that mean that my Tras setting of 0 is seriously messing things up?
I've tested it for a few hours in Memtest86+ (27 loops of #5, couple of full runs) and all seems stable .. 17 hours of Prime95 further proves that.

Logically, a setting of zero would allow pages to *not* be read anyhow.
These spoony 754 boards, a year on in and sometimes i still can't work out what half the options do, ehehehe

[Eldonko]

About a month ago, I tested all values of tras for the setup in my sig. I used SuperPI 1M with the anticheat mod. I did three runs of each tras value available and averaged the 3 times. The result was tras 5 was the best for my 3200el rev 2.

Outcomes will prolly be different for every type of MB and RAM, all you gotta do is test to see which is the best. :)

[craig588]

The Tras 11 thing was just a bug with the NF2 (Possibly 1 too, but I don't own one and havn't been able to find any testing with one) Every other chipset is better with lower values.

[EMC2]

It depends on the memory controller, the memory, and the BIOS... From the hardware side of things, most memory chips can be set to execute either 2, 4, or 8 word bursts (for example TCCD). Best tRas value depends on that burst length which the ram is set to (given that the controller can be setup for the same burst length). What the BIOS actually programs everything for is the software side of things.

Peace :toast:

{addition} Memory usage patterns of a benchie come into play as well... if the bench access patterns are exceedingly random, then low burst lengths help, if they tend to have localized access patterns, long burst lengths help.

[robberbaron]

tRas 12 and 13 offer the best latency for my CH-5 with a CAS of 1.5. tRas of 0 and 1 give about the worst performance.

[mushk1n]

ok so

2-2-2-11 > 2-2-2-5

but does that go for other timings...

2.5-3-3-11 > 2.5-3-3-7

and 3-4-4-11 > 3-4-4-8

? is this the case for everything?

[Mardok]

I try only 5-8-11 on my BH-5 and DFI NF4:

2-2-2-5 > 2-2-2-8 > 2-2-2-11

[Aguanteboca]

tRAS 0 is something I am researching about cause logically that makes no sense...you cant read a book without opening it first can you?

[[XC] 4X4N]

I have tested all tras settings 5-11. 6,7,8,9 are just about the same, but 6 gives me the best bandwidth (barely) in everest. Tref is the setting that seem to effect things the most.

[Mardok]

Test on BH-5 270MHz 2-2-2-x and DFI NF4 9x300 div180:

2-2-2-0
Sandra : 7669/7579
PI 1MB : 30.891s

2-2-2-5
Sandra : 7664/7577
PI 1MB : 31.140s

2-2-2-8
Sandra : 7662/7571
PI 1MB : 31.234s

2-2-2-11
Sandra : 7650/7558
PI 1MB : 31.703s

Then:

2-2-2-0 > 2-2-2-5 > 2-2-2-8 > 2-2-2-11

Screen's:

http://members.lycos.co.uk/m4rd0k/xs/2-2-2-0.png
http://members.lycos.co.uk/m4rd0k/xs/2-2-2-5.png
http://members.lycos.co.uk/m4rd0k/xs/2-2-2-8.png
http://members.lycos.co.uk/m4rd0k/xs/2-2-2-11.png

:)

[[timko]]

Nice dig on a 1yr old thread!

Can people please get it into their heads that this tRas 11 thing is for NForce 2 only.

So everyone else out there can stop wasting their time and benching their NF4 and NF3 systems trying to disprove the thread poster :)

[Mardok]

@[timko]:

Ok, but i want compare tRAS=0 to tRAS=5/8/11 not NF4 to NF2...