DDR2-533: Extra Latencies? So what?

[Mephistopheles]

For those thinking that DDR-2's extra latencies will matter, this is a great question: Why would it matter? THG showed that, with current chipsets, timings of 2-2-2-5 will yield performance differences that are minimal when compared to timings of 3-4-4-8. So going DDR2-533 should do only good, as far as we can tell now.

Unfortunately, it'll probably cost about twice the price of current DDR-1. But DDR, when first introduced, also costed twice the price of regular SDRAM... Prices will go down as more manufaturers compete and establish DDR2.

Wouldn't this mean that AMD's official statement was a little bogus?... Latencies won't hurt DDR2 that much... Problem for them, of course, is that they'd have to - guess what - probably change sockets again. Because of the on-die memory controller.

<font color=red><b>M</b></font color=red>ephistopheles

 

[TheRod]

I don't know if it technically possible to make a DDR/DDR2 controller on a single chip. If it's possible, AMD and chipset maker could make MB with DDR/DDR2 support.

I t will be great to compare DDR/DDR2 system performance. As you stated, maybe latency will not affect that much the performance of a system. But today, with speed increase of only a few percentage with each new CPU, every little difference in memory controller or RAM speed will make a difference.

P4C 3.2 to 3.4, would only be 6% increase in core speed.

--
Would you buy a GPS enabled soap bar?

 

[gothitbycar]

Don't you find it strange that THG did that article a few days after the fiasco with the A64 3400+ and the TakeMS memory? Im starting to find these reviews a bit suspicious. That review seemed like a "cover my butt" review.

Anyway, isn't there another advantage to DDR2 memory? I think it transfers 4 times per clock as opposed to 2. I'm not positive but I'm pretty sure.

-----------------------
Add witty comment here and beat this bad boy...
A64 3200+ w/ 1 GB DDR 400
2 36 gig raptors in Raid 0 complemented by an AIW 9800 Pro

 

[Mephistopheles]

That review seemed like a "cover my butt" review.

Maybe... Question is, did it cover their butt? I think it more or less did...

Anyway, isn't there another advantage to DDR2 memory? I think it transfers 4 times per clock as opposed to 2. I'm not positive but I'm pretty sure.

Yes, there is the advantage of better clock possibilities. DDR-2 will most likely be clocked at 533Mhz (2x266) when launched. But it comes with higher latencies, which might impar performance... or not, like THG's article suggests!

<i><speculation>
It is exactly this fact - that DDR-2 will be introduced at 533Mhz or 2 channels of 2x266 - that makes me think that maybe Alderwood and grantsdale will come with 1066Mhz FSB... It would give them one hell of an edge. Plus, they could launch two other prescotts to go with alderwood - the 13x266=3466Mhz and 14x266=3733Mhz. Think about it, compared to the A64, 3700+ the 3.733Ghz will feature 1066MT/s, compared to HT's 1000MT/s in AMD's platforms, and an extra frequency notch if ratings were to be compared. And a 3.73Ghz Prescott is probably quite enough to compete or even outperform the 3700+, if coupled with dual-DDR2-533 and alderwood. It's probably not too hard to do, because it doesn't add too much heat... The 3.73Ghz should be doable!
</speculation></i>

<font color=red><b>M</b></font color=red>ephistopheles

 

[Mephistopheles]

<i><b>(update)</b> Apparently, DDR-II modules of 2GB and at 533MT/s seem to have latency numbers of 3-4-5, which is to say that it is about the worst case scenario for DDR, according to THG (3-4-4). Bear in mind that this is mission-critical memory for servers, and at 2GB, so they should have conservative latencies! Kingston had less agressive timings (3-3-3 for PC2-3200 and only 4-4-4 for PC2-4300). At least one company has already managed to get DDR2 at 667MT/s operational.</i>

In any case, I'd expect the numbers to get better as memory manufacturers do their tricks. Hopefully.

<font color=red><b>M</b></font color=red>ephistopheles

 

[juin]

AMD is right.DDR2 internal timing might not be that bad but there internal clock speed is low like RDRAM.RDRAM or DDR-2 run at 133 mghz.So a CAS lantency of 2 equal 14 cycle on DDR-2/RDRAM and only 10 cycle on DDR1.If you want we can compare a 128 bit PC1600 vs a 64 bit Pc3200.Bandwith is equal but internal timing are not the same.If the ratio chipset lantency/ram lantency change for more ram lantency the intergrade controleur lose a bit of is important.For intel more bandwith is better that less lantency even maybe a 266 mghz MCH to have 1 dead cycle that should help a lot for memory read.Maybe AMD should think about a RDRAM controleur for there Opteron as there a lot of server chip that use RDRAM and will be happy to not lose those module.I feel that Jedec just want to produce a cheap not so bad performing dram.

I dont like french test

 

[juin]

Post deleted by juin

 

[Mephistopheles]

Oh, I didn't realize that DDR2 worked by quad-data-rate. This way, DDR2-533 is actually 4x133 and not 2x266?...

A little frustrating... I'll read up on DDR2. Does anyone know of a good article on it?

<i>edit: Wait a minute, DDR1 1 bit per cycle? That cannot be right! DDR2 2 bits per cycle? No... DDR1 at 200Mhz is DDR400! This would mean two bits per cycle...

And if DDR2 is unlike DDR1, which is to say it's 4 bits per cycle, then how come it is <b>D</b>DR and not <b>Q</b>DR? Something is not right here!</i>

<font color=red><b>M</b></font color=red>ephistopheles

 

[Vapor]

Since when is DDR-II quad data rate??

Maxtor disgraces the six letters that make Matrox.

 

[Mephistopheles]

It is <b>not</b>, as far as I know! But Juin seemed to indicate that DDR2-533 would work by transferring 4 bits with 133Mhz frequency, in his post... And that got me confused. I'm checking on it, but I'm still leaning towards 2 bits ber cycle...

<i>Edit: OK, what a mess, both are double data rate memory. 2 bits per cycle. Which makes me think that... I didn't understand Juin's post. Sorry there...</i>

<font color=red><b>M</b></font color=red>ephistopheles

 

[Vapor]

Yeah, it's not. DDR-3 isn't even QDR. All it is slightly restructured DDR-I to enable faster speeds, it seems. I'm also leaning towards 2 bits per cycle.

Maxtor disgraces the six letters that make Matrox.

 

[Vapor]

Okay, post-edit reply: Juin was wrong. Simple as that.

Maxtor disgraces the six letters that make Matrox.

 

[juin]

Little explanation between fetch size and data rate.

--------Fetch size--data rate--Io buffer--size--command
DDR----- 2 bit---DDR---1/1---64 bit---SDR
DDR-2---- 4 bit---DDR---2/1---64 bit---SDR
RDRAM---- 4 TRof2---DDR---8/1?--16 bit---DDR

So at 133 mghz DDR 1 send 2.1 gb/s
DDR 2 send 4.2 gb/s
RDRAM send 2.1 gb/s

DDR-2 have extra latency as it take a longer time to have 2 bit that just 1.

http://www.samsung.com/Products/Semiconductor/common/product_list.aspx?family_cd=DRM070202

DDR-2 667 ready with timing of 5-5-5
DDR-2 533 ready with timing of 4-4-4
DDR-2 400 ready with timing of 3-3-3

http://www.elpida.com/en/products/ddr2.html

same thing for elpida


I dont like french test<P ID="edit"><FONT SIZE=-1><EM>Edited by juin on 01/20/04 03:53 PM.</EM></FONT></P>

 

[Mephistopheles]

Valuable information there, juin. Thanks... but there is still one thing I don't quite understand. This site you mentioned said that DDR2-400 works at 200Mhz. And it is well-known that DDR2-400 transfers 3.2GB/s, which is in agreement with the name, "double data rate". But you said...

So at 133 mghz DDR 1 send 2.1 gb/s
DDR 2 send 4.2 gb/s
RDRAM send 2.1 gb/s

This cannot be true, because DDR2@133Mhz should only transfer about 2.1GB/s, same as 133Mhz DDR1. How could DDR2@200Mhz Transfer 3.2GB/s, and DDR2@133Mhz transfer 4.2GB/s? That wouldn't make any sense...

Samsung's site even says that DDR2-400 has the same transfer rate than DDR-400!...

I must be missing something here...

<font color=red><b>M</b></font color=red>ephistopheles

 

[Mephistopheles]

Can DDR-II reach 250Mhz? (Currently you guys said it's around 166MHz)

Yes. State-of-the-art memory modules of DDR2-667 have been produced already. These modules work with a frequency of 333Mhz, so as to result in 2x333=667. (we presume, here, that 333Mhz is 1000/3, so that 2x333Mhz is 2000/3 = 666.667, rounded up to 667Mhz)...

If it's a DDR550 vs. DDR-II 667, who will win? (Anything to do with timing as well)

This is a fair enough question... I can't answer that for sure. I think Samsung might. Maybe Infineon or Micron, too. But if I had to take a guess, I think I'd say that DDR2-667 will probably perform more respectably than DDR550; I'm the optimist who thinks that the added bandwidth will more than counter the added latencies.



<font color=red><b>M</b></font color=red>ephistopheles

 

[juin]

DDR-2 transfert twice as much with the same clock speed that DDR1 but have higher lantency AL (additif lantency as Jedcc call it.)

I dont like french test

 

[juin]

This is a fair enough question... I can't answer that for sure. I think Samsung might. Maybe Infineon or Micron, too. But if I had to take a guess, I think I'd say that DDR2-667 will probably perform more respectably than DDR550; I'm the optimist who thinks that the added bandwidth will more than counter the added latencies

For intel yes but for AMD that a others story.Where this will help is in server.In big I2 cluster only Pc2100 is use.Now they will be able to use DDR-2 533.

I dont like french test

 

[juin]

Yes original NV30 was using DDR-2 250 mghz.On desktop i will be very surprise

I dont like french test

 

[Mephistopheles]

DDR-2 transfert twice as much with the same clock speed that DDR1

Erm... My point was that it does not. According to samsung.

<font color=red><b>M</b></font color=red>ephistopheles

 

[Guest]

>If it's a DDR550 vs. DDR-II 667, who will win? (Anything to
>do with timing as well)

I read an interview with an infineon chap a while ago, who claimed DDRII-533 would roughly perform like DDR-400. Though much will depend on the apps, and even more on the cpu (I'd wager a bet the P4 core is generally more bandwith sensitive and less latency sensitive than a K7/8 core, wether it be because of better prefetch, other cache design or longer pipeline). So for the P4, DDR2 might give a small performance boost, while its probably neglectible for the K8.

Either way, I wouldnt expect big things from DDRII. Sounds like a small leap like going from FPRAM to EDO Ram. BTW, historically, no new memory type has given a major performance boost right away; not EDO, SDRAM, DDRRAM, or RDRAM. In the long term, and as cpu'd and fsb's scale further up, difference might get more tangible, but I wouldnt hold my breath and wait to buy the first DDR-II motherboard on the market.


= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[juin]

agreed

There might be a small decrease with K8 core.On intel side that wont help much before 1066 FSB that should be release in Q3 Q4

I dont like french test

 

[juin]

http://www.lostcircuits.com/memory/ddrii/2.shtml

I dont like french test

 

[Mephistopheles]

Oh, I get it... So, according to this article, DDR2-400's functional core works at an actual frequency of 100Mhz, and DDR-400's works at actual frequencies of 200Mhz...

In that case, DDR2 offers less heat dissipation, less voltage, and less expenses than current DDR. Sounds good. So think about it: in some terms, it's like going back to DDR-200... which would explain some of its disadvantages. However, once DDR2-800 comes out, then it will have completely made DDR obsolete: the core would be the same as DDR-400, but it would transfer twice as much... It's going one step forward and two backwards, but it also represents that the path will be clear in the future... 'cause DDR is hitting a "brick wall" that has to be sidestepped, I guess.

All this means that there's enough room there for an excellent price reduction for DDR2. Just let the manufacturers relearn their business...

Thanks for the info there, juin. I still have to make sense of how the facts I know interact, but I think I'm on my way...

<font color=red><b>M</b></font color=red>ephistopheles

 

[juin]

Most manufacturer dont have any willing to go with more expense memory neithers the market.Rambus story tell long.

I dont like french test

 

[juin]

Some manufacturer was implying that DDR2 might consume 75% less power that DDR.This will be welcome in server also.

My little theory is intel was largely involve in this like in DDR400 thay have made most of the standart.This new memory type fit better with P7 that K8 give them a little extra edge.On SMT chip bandwith is much more usefull that any thing else.

I dont like french test