Low latency, HIGH HTT, who wins??

[dexx]

From what i've read in this forum and others, i would guess that gaming and app performance would be comparable for the following speeds:
200Mhz 2-2-2-5 == 250MHz 2.5-3-3-7 == 300Mhz 3-4-4-8.

[Revv23]

i think 200 2-2-2 would be slower then 300 3-4-4...

anyone have anyting to prove that?

[burningrave101]

Raising the HTT speed doesn't do anything at all. The HTT speed has no impact on performance. Unless you lower it way down from 1000Mhz to like 400Mhz then its a pointless setting to even worry about overclocking. There is nothing that can make use of the kind of bandwidth the HyperTransport link can provide on today's systems so overclocking it wont improve performance. The HTT link is nothing like a FSB which has a memory controller that operates at the same frequency.

There are two main external links on an Athlon 64 platform. The HyperTransport link which goes to the chipset and the link between the CPU and the RAM. The link between the CPU and the RAM transfers data at the effective speed of the memory (i.e. 200Mhz default x 2 DDR). The integrated memory controller in the CPU operates at the same frequency as the CPU speed and it accesses the memory at that speed.

Timings play no real huge impact on performance ither. The difference in today's games from the tightest of timings to the loosest of timings is on average 3-5% at the most. That means if your getting 100 fps (which in most new games you wont be) your losing 3-5 fps, big deal.

If you want top PC performance in todays games and applications then save your money and buy 2 gig of cheap RAM instead of wasteing your money on RAM that costs $200+ for just a gig. In games and applications that can use more then 1GB of RAM you can definitely tell the difference in performance where as with the timings and speed of the RAM and HTT you couldn't possibly tell the difference.

http://www.hardforum.com/showthread.php?t=886780

[mdzcpa]

If you want top PC performance in todays games and applications then save your money and buy 2 gig of cheap RAM instead of wasteing your money on RAM that costs $200+ for just a gig.

Practical advice maybe...but certainly not Xtreme

[burningrave101]

Practical advice maybe...but certainly not Xtreme

The performance in a game that can make use of more then 1GB of RAM will definitely be Xtreme compared to the performance you'll get from only 1GB of overclocking RAM that costs twice as much .

I've done quite a bit of research into this at different times and i've had alot of expensive memory like the OCZ PC3200 Platinum Rev2 that i'm using now but i'm selling it and buying a couple gig of much cheaper RAM that will actually show a difference in performance.

If your benchmark crazy and thats all you do then RAM that will overclock well will improve your performance some but for any PC user that actually uses their machine for something productive or for gaming then its a waste of money.

[Revv23]

The performance in a game that can make use of more then 1GB of RAM will definitely be Xtreme compared to the performance you'll get from only 1GB of overclocking RAM that costs twice as much .

I've done quite a bit of research into this at different times and i've had alot of expensive memory like the OCZ PC3200 Platinum Rev2 that i'm using now but i'm selling it and buying a couple gig of much cheaper RAM that will actually show a difference in performance.

If your benchmark crazy and thats all you do then RAM that will overclock well will improve your performance some but for any PC user that actually uses their machine for something productive or for gaming then its a waste of money.

i beg to differ, not many games use more then 512 much less 1gb, so imo once you have a gb then a fast GB is faster then an OK 2gb.

[burningrave101]

i beg to differ, not many games use more then 512 much less 1gb, so imo once you have a gb then a fast GB is faster then an OK 2gb.

I think your behind the times if you still believe there aren't many current games that use more then 512MB of RAM.

http://www.hardforum.com/showthread....&highlight=RAM

http://www.hardforum.com/showthread....&highlight=RAM

Even older games like Elder Scrolls III: Morrowind that i play quite often performs MUCH better with 1GB of RAM over 512MB.

Most games will play fine with 1GB RAM but there are some that perform much better with 1.5-2GB of RAM and that number is going to continue to increase this year as new more advanced games are released.

Games can't make use of alot of bandwidth and with a dual channel memory controller on the s939 Athlon 64's overclocking the RAM isn't netting you anything and the HTT speed has zero effect so your left with timings playing a roll in your performance and i doubt anyone here can say they can tell the difference in a loss of 2-4fps in-game from more lax timings but i'm sure they could tell the difference in performance of games that use over 1GB of RAM when they have 1.5-2GB installed and their no longer hitting the page file.

Not to mention the fact you dont have a full gig of RAM in use anyways. I'm running XP and all i have open are a few IE windows and i'm already down to 712MB.

2GB of RAM will also greatly improve your performance in intensive applications that require alot of RAM and will allow you to multitask much better because your not having to use the swap file much.

[Revv23]

okay so DX9 engines and morrowind.

sure, i play with 1gb, and i know its faster then 512, but there really arent that many games that perform a ton better with one or the other, im sure 2gb will be a standard soon, but i dont think its that time yet.

[Clevor]

I found out my PQI TCCD can do 300, 2.5-3-3-6 at 1T on the DFI NF4. That's 2x512.

From the standpoint of all four Marks and Aquamark, I have seen results at 300x9, 2.5-3-3-6 versus 260x10, 2-2-2-8, but I'm at work now so don't quote me on anything. Aquamark was equal or better on TCCD. 3DMark2000 was also surprisingly high, but BH-5 was a tad better. In 3DMark2001, the nod definitely goes to BH-5. 3DMark2003/2005 is pretty much video card dependent, but the nod went to BH-5.

In SuperPI 1M, hit 30 sec on BH-5 and 31 on TCCD. I have since found the PQI maxes at 308x9. I think here, all benches are neck-and-neck.

But ya gotta like 4012/4180 unbuffered with the TCCD at 300x9.

Keep in mind all my results are with 2x512. 2x256? I don't go there. You get much better unbuffered with 2x512 on A64, to the tune of 3743/3821 (I forget) at just 276x10, 2.5-3-3-6. And yup: TCCD+high fsb+big multi is hard to stabilize from a 3D standpoint.

I've had good luck with old BH-5 and new UTT, and personally feel it's easier getting stuff to do 260, 2-2-2-8 than find 2x512 TCCD that can actually, truly bench 3D at 300, 2.5-3-3-6. 2.5-4-3-6 is a different matter, as I found almost all the TCCD I have do well over 300 at 2.5-4-3-6 on the DFI board.

I plan on doing more benches later.

[Spajky]

Hello Xtremers!

I have read all 5 pages of this thread & I have some comments to add:

-------------------
1st site:

@ bachus_anonym : you took the right test to do benchmarking (WinRAR´s built in), but you really are not sure 4 what that test serves! (but you are close!)

@ charlie : Stability= The ability to run reliably during a required task .. IMHO NOT!
Stability=the ability to run reliably ALL different tasks @ hottest summer day for hours!

@ uwackme : probably higher mem.clock with higher latencies gives better results than both lower ones if frequencies of ram are enough far away from each other (like 250 & 300MHz clock)

@ HARDCORECLOCKER : forget Sandra mem.bench, it has almost nothing to do with real life performance of ram subsystem

@ Dissolved : IMHO better running memory at lower latencies @ 300MHz than 250MHz & tightened latencies if the CPU speed is same!

@ macci : IMHO better to use lower mem.divider if memory can handle it (higher real mem.clock)

@ bachus_anonym : >WinRAR benefits HUUUGELY from tweaked BH-5< :-))))) you took the right test to do benchmarking again! (stick with it! - you will find out later why!)
>The point of this comparison is to show what numbers I get at same CPU speed... Not getting as high clocks as possible...< Very wise way of thinking & using WinRAR!


2nd site:

@ saaya : as I stated also before, I could agree with your thinking about stability ... & FSB is the bus between mem.controller (not memory exactly, can run at different speed!) & CPU.
>i really would like to make a scale that shows where tccd and bh5/dh5 meet!< he he, use WinRAR´s built_in test & do your experiments for your setup!
>expirience has always been that if an oc isnt prime stable then games will crash sooner or later< .. sure right! ...

@ Holst : >the difficulty in comparing one to the other is ...< is not if using that WinRAR´s built_in test ...

@ esdee : >it's all about how high your ram can get< Right, IMHO ...


3rd site:

@ iddqd : >Otherwise, high bandwidth gains very little< True IMHO in real life ..

@ Kunaak : >I am however interested in seeing other peoples results< you will, you will, but later results with WinRAR´s built_in test..

@ macci : >interesting stuff< I wonder what you will say after you will read my article later ..
>But why would you want to run different memory sticks w/ the same (un-optimal) timings?
Wouldn't it be best to run em all as fast as they go (as far as the settings go) and then do the tests< I agree too also with this, Right question & testing with WinRAR´s built_in test !

@ bachus_anonym : >the comparison wouldn't be fair< it would be if if you squeeze them both to max. & than recalculate WinRAR performance per real clock, so you know how much one is better than other... (like I did as you will find later!)
>I'll be doing another comparison very soon< read before my article!

@ uwackme : >What MATTERS is how fast the MEMORY bus is.... the bus between the memory controller and the sticks of ram< totally correct!
>The answer is.... the most MB/s is what is best, how you get there is IRRELEVENT< totally correct again!
>You guys with A64 systems and good VX and TCCD sticks can answer the question by simply measuring the MB/s the various configurations result in.< Yes, with using WinRAR´s built_in test! & I agree also with all you described in that post! You are really a clever mind!

@ Rabbi_NZ : >aaah yeah.. what do you mean by MB/s? How would you go bout measuring that? Cos Sisoft Sandra sure as heill dont tell me jack bout any systems performance.... it's great for overclocking competitions and pretty screen shots, but for anything "real" it means diddly squat< You are right too & you will soon understand why I stick so much mentioning WinRAR´s built_in test after later reading my article!


4th site:

@ uwackme : >on the mB/s, the Sandra's is a guesstimate. you could use a Oscilliscope and memtest to actually SEE and measure for real what the MB/s transfer rate really is< No need of that, thats why we have the newest WinRAR´s built_in test! (IMHO now if reading this, you are already sick of my mention it again!)

@ macci : >In theory it would seem so but in reality ('real world performance') CAS latency doesn't really make any difference at all (on modern platforms).< Yes I also realized that too time ago that with higher & higher mem.clock, latencies matter less & less (if not to big) ..

@ Carlos Henrique : yes when mem.clocks are not far away is better to use them in sync with Fsb & a bit tightened latencies ...

@ macci : BTW, Happy Birthday 2 U !
>ASync DOES mean a hit in perf per memory clock< Yes but se above (Carlos)

@ Dissolved : >what i would compare is Avg memory overclocks< Thats why i propose reading later my article!

@ xsky : >someone said that memtest screens were good for comparison< No, IMHO are as bad as Sandra results!


5th site:

@ saaya : >i HOPE that this discussion will go on< LOL, MAYBE I´m doing it right now!

@ flexy : >the point is that you, as a human, can NEVER really evaluate what is "reliable" since you might just not notice when the CPU calculates wrong due to too high a overclock<
Yes if is overclocked over "sweet spot" (see down too)
>that bandwidth is the only thing which counts< uwackme IMHO said right contrary ..

@ burningrave101 : >The integrated memory controller in the CPU operates at the same frequency as the CPU speed and it accesses the memory at that speed< not completely true, you forgot about mem dividers & that if the core has non integer nultiplier (eg.10,5), Ram never works in_sysnc with Fsb!
>Games can't make use of alot of bandwidth< but need a lot of true speed of moving mostly compressed textures data thru mem.subsystem & NB to AGP port & videoCard and if that is slow, there is one of reasons of game stuttering IMHO ...

--------------------------

Sorry If in this my post was sometimes not properly quoted before comments, but I was reading & writing it offLine (I´m on DialUp!).

I have also something to add further: when OC-ing memory using much higher voltage on mem.modules than stock ones CAN impact on real speed of them, so pushing voltage up too much, can be contra-productive, even if some benches do not show that correctly! The same is for other stuff like CPU. The point is to find a so called "sweet spot" when on all cumulative settings, the machine in real life is fastest & most responsive! Can be lower few MHz than max the hardware can handle stably! A lot of times pushing stuff higher does not bring any real gain (just additional heat!) in every-days use of a machine. Well, searching for certain benchmarking records in another job ... ..


That mine article I wrote time ago & is my contribution now to this thread started by bachus_anonym. Read it, think about my opinions & "sleep it over"; maybe next day you will get the another idea how to benchmark stuff alternative way! :-)

PS. Some reviewers from different hardware sites time ago used that WinRAR´s built_in test for wrong purpose (benching CPUs capability or compressing data-wrong!), so I explained them that /xbitlabs, anandtech etc./ by mail.

SuperPi is a great benchmark capable of showing real life performance of a system, but not only for Ram or Cpu separately, because depends cumulatively on Cpu raw calculation power, FSB, L2 size & mem.subsystem performance all together in one. So for example with all other parts & settings & frequencies same but with same clock/Fsb CPU but L2 different, you will not IMHO get same final result.

--
Regards , SPAJKY ®
3rd Ann.: - "Tualatin OC-ed / BX-Slot1 / inaudible setup!"

EDIT: sorry, no advertising...
-saaya

[Spajky]

@ saaya : Damn You were fast! Advertising? What advertising! Sorry, thats was my Home Page (non-commercial!) in my signature added where you can read that mem.benching article I mentioned! Oh, doesn´t matter anyway... (first time in 2y that that link posted was cancelled somewhere & I posted on a lot of different Forums worldwide)

@ OTHERS: Well, now is under my profile linked (like saaya´s HP same) or simply copy/paste my nick into Google; You´ll get there IF U R enough curious to read it ... Bye ...

[saaya]

sorry for this kind of welcome ^^

it doesnt matter if its a comercial site or not, if you link to an article thats always ok
oh and your site looks messy in firefox, can you fix it please?

[Spajky]

Redesign of a page will soon or late come as I planned, but I still have too much interesting stuff semi-prepaired to finish first to put it there before, to fill "gaps" on page...


The "about BENCH´ING MEMORY -real life" (my research) article ...

[saaya]

hmmm all the weird symbols are confusing me, and some are even moving
but im going to read it later

[Gogar]

Does anybody ever realise that more relaxed timings at a higher HTT might actually be the same latency in absolute terms? Or am i smoking crack here?
The memory timing numbers are clock pulses right?.. If the HTT increases the time each clock pulse takes gets shorter.

[wwwww]

Does anybody ever realise that more relaxed timings at a higher HTT might actually be the same latency in absolute terms? Or am i smoking crack here?
The memory timing numbers are clock pulses right?.. If the HTT increases the time each clock pulse takes gets shorter.

You're spot on - from my experience, in general, or at least superpi, htt speed makes no difference whatsoever - you aren't going to use that bandwidth anyway.

However, Increasing the htt also increases timings. eg. Cas 2.5 @ 250MHz is a 10ns cas rating. Cas 2.0 @ 200MHz is also 10ns hence being the same.

[[XC] 4X4N]

I have been benching my venice a bunch in the last few days. Here are two shots at 10x280. One with twinmoss ch-utt at 234, 2-2-2-6 and one with ocz plat. rev 2 tccd, 280, 2.5-3-3-8. Even though the tccd shows more bandwidth, the utt is faster in super-pi and sandra cpu benchmarks, and thats at almost 50mhz lower! So I would say that the tighter timings at lower mhz will win every time.