Low latency, HIGH HTT, who wins??

[HARDCORECLOCKER]

An interresting discovery I made the other day macci:

252 x 10 1:1 > 280 x 10 /11 by 200MB/s in unbuff .

ASync DOES mean a hit in perf per memory clock.

I always thought on A64 it only mattered what the actual mem MHz was.

- is it only me not understanding this ???

[Carlos Henrique]

- is it only me not understanding this ???

252 x 10(cpu multi) at 1/1(mem divider = 10) u got 252mhz mem clock

280 x 10(cpu multi) plus 11 mem divider u got 254mhz mem clock, but 252x10 = better by 200mbps

[HARDCORECLOCKER]

252 x 10(cpu multi) at 1/1(mem divider = 10) u got 252mhz mem clock

280 x 10(cpu multi) plus 11 mem divider u got 254mhz mem clock, but 252x10 = better by 200mbps

THX - my bad, didn't understand "11 mem divider" bacause thought of 133-166-200..........

[macci]

An interresting discovery I made the other day macci:

252 x 10 1:1 > 280 x 10 /11 by 200MB/s in unbuff.

ASync DOES mean a hit in perf per memory clock.

I always thought on A64 it only mattered what the actual mem MHz was.

What speeds were you actually comparing there? 252x10 /10 vs. 280x10 / 11 ? or 252 x10 /10 vs. 280x9 /10?
Did you try any other tests? (like SPi 8M for instance?)
Sandra may not always show correct results if one is changing multipliers from windows.
Neither of those speeds is actually '1:1'. They are both divided from the overall CPU freq. and as long as the mem freq. and CPU freq. remains identical overall performance wont chance either (believe me I've tested this time and time again w/ both s754 and s939 stuff).

[Zeus]

Copied and pasted from this thread.

Testbed:
Mobo: DFI LanParty UT NF3 250GB
CPU: 3400+ NewCastle s754 (watercooled)
PSU: Enermax 465watt 3.3V @ 3.5V
RAM: 1x512mb OCZ EL PC3200 rev2 TCCD/ 1x256mb Kingston HyperX BH-5
OS: Windows 2000 SP2

I have tried to keep the CPU speed the same in all benches ~2860MHz.

Let's start with the BH-5, i have to admit this is only one stick of 256mb though.

PiFast @ 11x260 1,5-2-2-5 1T:

As you can see, the timings are pretty tight.

SuperPi 1M @ 11x260 1,5-2-2-5 1T:

Note the 31.578 seconds in process timer.

Now the TCCD, this is one stick of 512mb.

PiFast @ 317,8x9 2,5-3-3-7 1T:


SuperPi 1M @ 316,5x9 2,5-3-3-7 1T:


I know not 100% fair tests but it gives some idea.
I guess TCCD at 2,5-3-3-7 needs about 30MHz RAMspeed to keep up with BH-5 at 2-2-2-5.

[Rabbi_NZ]

I guess TCCD at 2,5-3-3-7 needs about 30MHz RAMspeed to keep up with BH-5 at 2-2-2-5.

But your difference is 56mhz... and you had 256mb more so Bank Interleaving may have come into effect there.

Where did you get the 30mhz figure from?

[bias_hjorth]

Well the tccd are a tad faster on the above screenie so a cut down on mhz would probably ad up in the end. Although I would say aprox. 40mhz to keep up. Which pretty much matches 300 2.5-3-3-7 vs 260 2-2-2-5

[Zeus]

But your difference is 56mhz... and you had 256mb more so Bank Interleaving may have come into effect there.

Where did you get the 30mhz figure from?

I know, i know, like i said, it's not 100% fair but it gives some idea.
Bank Interleave wasn't enabled, because i tested with an old bios that didn't have the BI option.

30MHz is a guess based on what i've seen around here.

[RAMMIE]

And we know it's alot easier and cheaper to get 260 2-2-2-5 than 300 2.5-3-3-7.

HUH? Not where I come from.

[bias_hjorth]

I have to agree with Rammie.. It most certainly depends on where / what you shop..

[isp]

There MUST be a point that TCCD at 2.5-3-3 PASSES BH5/VX at 2-2-2...

We still trying to figure this out huh? I don't have a set of each to play with so I'm sorry I can't be of any info. I'm sure some guys who have done extensive testing can speculate what it would take to beat really fast 2,2,2

I know Macci said earlier in the thread he doesn't believe tccd can reach that speed needed but what is the speed we are thinking of atleast? Ballpark it?

[Dissolved]

one thing i think most people overlook is how far can both bh-5 and tccd scale?

bh-5 has shown to scale up to 283mhz 2-2-2-5
tccd has shown to scale around 300-320ish 2.5-3-3-7

there has to be some point where tccd might overcome bh-5 if fully tweaked.

what i would compare is Avg memory overclocks, not everyone can get 250+fsn on bh-5 and not many people can get 300+ on tccd. it seems the dfi nf4 mobo increases total fsb gained on memory then nf3 or other motherboards as well.

Im no pro, But i think tccd would be better to fully max out a cpu overlock as with mem div you have Alot of options to play with. bh-5 only goes so high, but the lower total memory mhz might hurt maxing out the cpu overclock unless you say you have a fx cpu.

I dont know what id want. Personally, id take the tccd over the bh-5 since it needs 3.6v+ to max out its full mhz. tccd really only needs 2.8 in most cases.
Plus 300+fsb looks a hellovalot cooler in screenshots :P

[mcnbns]

bh-5 has shown to scale up to 283mhz 2-2-2-5
tccd has shown to scale around 300-320ish 2.5-3-3-7

Not too many people have BH-5 that will get to 283, but lots of people have TCCD that will get to 300+ (althought maybe not at 2.5-3-3-7).

[xsky]

someone said that memtest screens were good for comparison:





These aren't my Screenshots...hope they won't be deleted too soon.

I have more than 4GB/s with Intel....is this because of the messurement?
Thought AMD was bandwidth king

[esdee]

there is one more thing... TCCD is out of production and bh5 have 'a revival'

I'd stick to Winbond chips
If there is no other that can do that i can do 2x256 BH5@270Mhz (10x270) Vs 2x256 Gskill (300x9).

Just tell what you want me to run (sandra/spi/pifast/everest)

[Napoleonic]

what about UTT?

[esdee]

what about UTT?

my UTT are 2x512 so bank interleave will outperform everything in 270x10

[Troll]

bh-5 has shown to scale up to 283mhz 2-2-2-5
tccd has shown to scale around 300-320ish 2.5-3-3-7

If I look at the BH5 stepping database, then Ill find, that 283 is really an extreme value.

So you have to take an extreme value for TCCD too and that's around 340-360!





And I think, that TCCD with this frequency will surpass BH5, even with relaxed 2.5 timings

[funkflix]

Mhm, don't think so!



108MHz difference and more then 2 seconds faster!

[bachus_anonym]

@Troll

This was not really best example... That screenshot is MrIcee's attempt at highest DDR clock but unfortunately, he can't tighten up those timmings...

360 8-5-6-2.5-1T is not really best "performance-wise", believe me... I'm getting 5s better SuperPi 16M time @ 345x7=2415MHz 7-3-4-2.5-1T.... That is with over 105MHz less on the CPU

On another note, my SuperPi 16M run @ 339x8=2712Mhz 7-3-4-2.5-1T is slightly better that 270x10 6-2-2-2.0-1T, about 5s to be precise...

So you see, at least in SuperPI benchmark, BH-5 is still pretty hard to beat...

[Troll]

You are true... that's why I wrote 340-360MHz.

bachus: no pifast times for your 345-2.5-4-3-7-1T available to beat gravdigga's 46.74s?

[derubermensch1]

Hi, I'm new to these forums and have been reading around trying to decide what memory I want for me new system. I'm curious as to if anyone tested this high speed/lax timing, medium speed/tight timing theory with Doom 3 yet as was mentioned in the original post? Curious as to which performs most reliably/best? I'm primarily into gaming : ) Was there ever a conclusion drawn for this or was this discussion moved to a different thread? (it seemed to just end for no reason) I'm swaying towards g.skill as it it can also do 2-2-2-5 in case the memory controllers on the coming dual-cores aren't as forgiving (don't want to waste $300 on memory), but if there's a discernable difference between TCCD at higher speeds and BH-5 at lower speeds in favor of the BH-5 then I might reconsider. Thanks!

[saaya]

uwackme, what the bus is capable of in theory is something different than where it already starts slightly limiting the performence...

no card needs more than 2 pciE lanes IN THEORY, but in practise we see that going from 8 to 4 lanes already shows a very small, but reproduceable performence difference.

same for the htt bus, maxing it out is one thing, but when it comes to getting the best performence possible increasing the htt bus has shown to have a minimal impact on some benchmarks, even though the bus isnt maxed out even at stock in theory.

michal, i cant wait for the tweaked bh5 vs tccd results to make a scale with the results and see where bh5 and tccd meet
thx a lot for the tests you ran so far and the time to sacrificed for it

Is it typical that Geil ultra X can run around 250-260mhz with 2-2-2-x timings at around 3v?

absolutely not! anybody is welcome to prove me wrong, and be glad if somebody did it, but those results are just incredible. make a thread with some more shots and results kunaak! thats really worth some showing off

and last but not least, great to see you participate in a discussion again macci!

Hi, I'm new to these forums and have been reading around trying to decide what memory I want for me new system. I'm curious as to if anyone tested this high speed/lax timing, medium speed/tight timing theory with Doom 3 yet as was mentioned in the original post? Curious as to which performs most reliably/best? I'm primarily into gaming : ) Was there ever a conclusion drawn for this or was this discussion moved to a different thread? (it seemed to just end for no reason) I'm swaying towards g.skill as it it can also do 2-2-2-5 in case the memory controllers on the coming dual-cores aren't as forgiving (don't want to waste $300 on memory), but if there's a discernable difference between TCCD at higher speeds and BH-5 at lower speeds in favor of the BH-5 then I might reconsider. Thanks!

i HOPE that this discussion will go on, lol

about your doom3/game question, check this:
http://www.xtremesystems.org/forums/...ad.php?t=48634

he tested doom3 and couterstrike with different latencies on his a64

[flexy]

Stability= The ability to run reliably during a required task.

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

Example is distributed computing, FFT calculations, Seti, Folding etc.....you as the person in front of the computer might see your PC running fine because it "doesnt crash"....while in reality the CPU has a hard time caclulating the right values....and especially with distributed computing this is "dangerous" since many wannabe-overclockers submut results which are just plain wrong.

In programs like P95, super pi etc. its obvious since these programs check their results against values....but you will never know whether your game/application "internally" calculates right since you dont get feedback.

[flexy]

Again, back to the original question.... HTTvs blah blah blah
[...]

very nice read !

I myself (and probably with me most here ) are constantly trying to push their h/w, mem sticks etc. on the edge....and i find me constantly on the edge of "what is JUST stable"....eg. if its an odd day and it rains i get "odd" problems and another day not. Becauase i try to push my TCCD with the tightest possible timings at 289 the same time pushing my CPU.

I should really investigate into the advantage of running 10x260 instead (of 9x289)....if its the case, as you said, that bandwidth is the only thing which counts.