DDR2-667/1066Mhz FSB preview

Toggle sidebar Toggle sidebar
Mephistopheles

Mephistopheles

Distinguished
Feb 10, 2003
2,444
0
19,780
...babelfished <A HREF="http://babelfish.altavista.com/babelfish/trurl_pagecontent?url=http%3A%2F%2Fwww.x86-secret.com%2Fpopups%2Farticleswindow.php%3Fid%3D107&lp=fr_en" target="_new">straight from x86-secret</A>... Seems 1066Mhz FSB/DDR2-667 Combo is considerably richer performance-wise than 800Mhz FSB/DDR2-533... If only this didn't carry EE price tag...

<P ID="edit"><FONT SIZE=-1><EM>Edited by Mephistopheles on 08/16/04 11:24 AM.</EM></FONT></P>
 
That's ok though. Not too many people are dumb enough to buy it at EE prices, but the good thing is that we see that intel seems to be heading in a good way in the near future. I'm too lazy to bablefish it, but if it's showing considerable performance gain than i'm happy indeed.

I still hope dual core comes a bit sooner. 😀 Just seems neat.

<A HREF="http://arc.aquamark3.com/arc/arc_view.php?run=277124623" target="_new">http://arc.aquamark3.com/arc/arc_view.php?run=277124623</A>
46,510 , movin on up. 48k new goal. Maybe not.. :/
 
such a long article, and the most interesting thing remains untested: how does DDR2-667 compare to DDR1-400 (or 500) in real world apps ? Of course it gives better bandwith in synthetic apps, but does it offset its latency handicap ? (note that even DDR2-667 offers considerable worse absolute and relative latencies as both DDR1-400 and 500, at least with the timings achievable today).

*sigh*, why is it every review out there always seems to forget to benchmark the really interesting things.. (like EM64T 32<>64)

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
(agreed on the EM64T 64-bit issue)

Sorry, I pointed to a non-babelfished link previously. I intended to link to an english translation right away; the offending link has been punished by correction.
 
Yes Juin, you are quite right, as usual. After all 2-2-2-5 is slower than 4-4-4-12 (relative latency) and 69ns is also slower than 76ns (absolute latency). But hey, maybe you didn't quite understand the article, as it was in French.. ?

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
*sigh*, why is it every review out there always seems to forget to benchmark the really interesting things.. (like EM64T 32<>64)
Click to expand...

I am sure you will not make that mistake as you are now in a position to improve on that (PCGamer), right?


BigMac

<A HREF="http://www.p3int.com/product_center_NWO_The_Story.asp" target="_new">New World Order</A>
 
.. but then, I'm sure I'd make 200 other ones, maybe more crucial ones if I'd start setup up the systems, and selecting/running the benchmarks. Remember, I'm a project manager, I'm good at telling other people what to do, not in doing things myself :)

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
BTW, it is <A HREF="http://www.pcgameplay.be/" target="_new"> PC Gameplay </A>

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
In fact they have the same internal timing you should know that.The number is base on best case lantency on a very small data i guess 64 bit.In real life more bandwith will give you better overall lantency and worst case scenario lantency will decrease.On a large block of random access DDR2 667 will have better latency.

i need to change useur name.
 
>In fact they have the same internal timing you should know
>that

I beg your pardon ?

>The number is base on best case lantency on a very small
>data i guess 64 bit

Here is a link for ya:

<A HREF="http://www.digit-life.com/articles2/ddr2-rmma/ddr2-rmma.html" target="_new">http://www.digit-life.com/articles2/ddr2-rmma/ddr2-rmma.html</A>

<b>As you can see, DDR2-533 has higher latency than DDR-400. </b>However, it shouldn't come as a surprise after we described the theoretical basis of the new DDR2 standard.

The difference between DDR and DDR2 latencies is practically invisible (3 ns) at a standard 64-byte memory read when the hardware prefetcher is active. But it becomes much more obvious in the case of the dual-line 128-byte read.
Click to expand...

I'm sorry, did you have a point ?

>In real life more bandwith will give you better overall
>lantency and worst case scenario lantency will decrease.

How so ??? No you're confused at best, higher frequency will give you lower (absolute) latency, but DDR2 relative latency is so much worse, it is NOT offset by the higher clockfrequency (unless you start comparing DDR2-667 to DDR1-333 CAS 3 or something). Read the tests for crying out loud, latency is higher, period. be it @533 versus 400 or 667 versus 500 (heck, even 400), be it 64 or 128 bit reads.

>On a large block of random access DDR2 667 will have better
>latency.

Read the frigging tests.

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
Hm, I'd expect that, if and when JEDEC standard latencies for DDR2-800 are disregarded in favour of a low-latency version of DDR2-800, we'd get better overall figures than DDR400. But this is still at the very least a year off, because the DDR2 process is completely new right now.

DDR400 is a powerhouse. It was initially believed that it wouldn't catch on and that DDR2 would be required, but DDR400 proved not only to be possible, but it's turned into the pinacle of DDR development, a real trump card.

It's a very good thing that memory manufacturers are, in fact, pushing DDR2-667 out the door much faster than expected. After all, it's the first exciting DDR2 there is; they shouldn't even have considered DDR2-400, for instance. DDR2-400 is an obvious loser. Which was also a lousy choice for Xeon servers, they should have tried a full DDR2-533 at least. They should even preferably put a DDR2-667 controller with the newest Xeon chipset (or better yet, put it directly on the CPU... or buy an Opteron :smile: )

I mean, DDR2 is theoretically much more easier to ramp in clock speeds. Why don't they jump at the chance to do so? DDR2-667 is already available.

<b>The early adoption of DDR2 (and the eventual quick speed bumps to follow for DDR2-800) is still no excuse to keep that outdated bus architecture.</b> DDR2 can easily be clocked up and timings could be made more agressive. Intel could move to something better, and one or another rumor has popped up about they conjuring something better up. Just hope they're true.<P ID="edit"><FONT SIZE=-1><EM>Edited by Mephistopheles on 08/16/04 04:39 PM.</EM></FONT></P>
 
<A HREF="http://www.xtremesystems.org/forums/showthread.php?s=&threadid=40155" target="_new">I ran some DDR2 tests</A>

Latency is everything with DDR2, so if you are running at stock speed you will want fastest possible timings. To reach DDR667 with current processors you will need to increase FSB to 250Mhz and set memory ratio to 3:4

For those who dont know the math. 250 / 3 * 4 = Mhz * 2

Asus claims to have a 2:3 ratio on the P5AD2 and with the newest beta bios its still not present/working.

(just picked the end post to reply, not directing at you)

<b>Kanavit owns you, get over it already!<b>
 
BTW, P4Man, that link of yours explained quite exactly why the first incarnations of DDR2 (400 and 533) are useless anyway; higher bandwidth which doesn't get used and higher latencies which worsen performance. Nice link. :smile:
 
I mostly agree. few things though:
Its not because DDR2 has higher latency (both read and write, and no matter what Juin claims) that its useless... it does offer higher bandwith still :) Furthermore P4s prefetching and HTT go a long way in hiding memory latency, so for P4 DDR2 shouldn't be too bad (and according to benchmarks, it isnt, it just isn't really any better yet either). Same goes for Xeon; surely DDR2 400 is better than the DDR1-266 (even single channel, or not ?) it was limited to until now. Also for Nocona especially, lower powerconsumption of DDR2 is also worth something

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
Wow, you managed 3-2-2-12 with DDR2-533? Doesn't look bad at all. Isn't the CAS Latency limited to being equal or greater than 3 in DDR2?

Well, 3-2-2-12 is good for DDR2-533; if in a year we can have higher-clocked modules <i>which can keep good timings</i>... Until then, though, DDR2 = nonsense for me.
 
Hm... right, maybe you're right, lower power consumption is important for Xeon (even more so if you consider the Nocona processors run so hot)...

You do have a very good point; DDR2 is just not better than DDR400 yet. And I forgot that Xeon never got access to DDR400... Which makes DDR2-400 a lot more interesting.
 
Math is not you strong point neither the technologie you are talking about.

DDR2 timing come from the point of view of the chipset wich only see the IO buffer.DDR2 667 and DDR 333 have the same timing Cas 4 from the IO buffer and CAS 2 in the cell array.You own link show that it the same timing for the cell array.



How so ??? No you're confused at best, higher frequency will give you lower (absolute) latency, but DDR2 relative latency is so much worse, it is NOT offset by the higher clockfrequency (unless you start comparing DDR2-667 to DDR1-333 CAS 3 or something). Read the tests for crying out loud, latency is higher, period. be it @533 versus 400 or 667 versus 500 (heck, even 400), be it 64 or 128 bit reads.

Reply:

Single stride lantency will be better on DDR 400 but large block will be better on DDR-2
A 1KB request will take less time on DDR2 667 that DDR1 400 or even 500.

i need to change useur name.
 
> 1KB request will take less time on DDR2 667 that DDR1 400
> or even 500.

And a 1 GB request might be faster on DDR2 533 9-12-12-36 than on DDR1-500 2-2-2-5. Does that tell us anything on LATENCY ? Let alone performance ? I thought not. I don't know *what* you measure it in, Gb/s, meters, decibel, dollar or ampere, but if you measure access latency in <b> nanoseconds </b>(between read and data out), DDR2 is slower than DDR1 at equal or slightly higher speeds. Both in theory and in reality, both in reading and writing, for either 64 or 128 bit blocks.

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
I would say that there should be a balanced mix of low latency and high bandwidth. Things being what they are, the best mix is currently still being offered by DDR400, and only DDR2-667 might offset that, and only if coupled with 1066Mhz FSB.

After all, a more adequate measure of memory system performance would be the sum of latency with the time it takes to transfer a data package of the average size that that particular CPU architecture demands. Therefore, if memory access requests come bundled, your average data package will be bigger, which will mask high latencies; if memory access requests are fragmented, latencies will mean a lot.

That's my logic, anyway, but I don't see any other way of painting the overall picture.
 
Those tests on Centon DDR2 were run at 3-2-2-5 1T, the memory will also run at 3-2-2-4 = faster than any current DDR1 @ 200Mhz

<b>Kanavit owns you, get over it already!<b>
 
I saw the thread, but I'm kinda lazy/at a loss what it tells us.. are you seeing better application performance (not just bandwith in Sandra) with those LL DDR2 modules as opposed to DDR1 ? If so, by how much ?

One thing though, your sig stated you used a P4EE ? it would IMO be more interesting to test with a regular P4, as the big cache of the EE goes a long way hiding latency and speed of the memory subsystem as a whole. Its quite possible the EE would benefit more from DDR2 as a normal P4 (and maybe even prescott over northwood). If you're bored, and have the required cpu's, feel free to test and post 😀

= The views stated herein are my personal views, and not necessarily the views of my wife. =
 
well im sitll not convinced that ddr2 wil stay around that long.

with fb-dimm on the horizone and ddr3 aimed to come next year, what is there to keep ddr2 as the mainstream choice when even by that time next year ddr2 wont be in the same price league as ddr at the ddr2 667 levels. ddr3 is just ocming too fast for ddr2 to get intrenched if it is really coming next year. all those rumors of amd skipping ddr2 all together arent so far fetched considering thier other options. it probably owuld hurt intle if amd skipped it all togehter and jumped right into ddr3, but not as bad as rambus did. the thing to wait for is to see if memeory manufactures really start ramping up ddr2 production form high end to value.
 
You must log in or register to reply here.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom