DDR vs. RDRAM at Computex:

[Novgrod]

http://www.theregister.co.uk/content/3/19452.html

If somebody wants to linkify that, it'd be great.

So what does everybody think? Will Intel cripple its i845 to help its i850 and Rambus, only to be upstaged by via?

 

[fkloster]

<< Will Intel cripple its i845... >>



i845 never had a chance to begin with. The P4 was not desinged with DDR in mind. Memory timing issues coupled with lack of bandwidth will be the 'nail in the coffin' for i845.

 

[shathal]

Mate - this is bound to be the benchmarking game all over again.

If puahed, I could probably end up configuring a system where a Celeron would beat an Athlon that
had 500 MHz or more (well - ok, that'd be pushing it).

A few BIOS change here &amp; there, a few drivers configured differently (or not loaded at all),
it's all that, I would say. Benchmarkin by PR - they're hardly fair. I'd be more inclined to
believe those numbers if Anand or Tom Pabst would bring them up.

Right - let's disregard the numbers then for a second.

As such, I've nothing got against DDR-chipsets for P4 - quite to the contrary, making a P4
system more financially available (i.e.: cheaper) makes good sense.

So - why isn't Intel willing to have VIA that license? I doubt it's got ANYTHING to do with
Inte's own chipset for DDR which will only come out next year - I think it's more likely that
they distrust VIA's chipset reputation (which - let's face it - has take a few nasty blows). And since the P4 is still a &quot;new&quot; product, they want it to run on a stable chipset. If VIA's chipset should be stable (would be nice ), then that would be good for VIA and Intel.

But I seriously distrust these numbers. It's like the benchmarks on AMD's site - if you have a look at the small-print you notice a few &quot;unfair&quot; settings. And then there's the things you can do in the BIOS as well, etc, etc.

The benchmark-game is just that - a game. Anyway - that's my 2p on the subject.

 

[ST4RCUTTER]

Linkified

 

[shathal]

Memory timing issues coupled with lack of bandwidth will be the 'nail in the coffin' for i845.

BTW - the i845 &quot;as is&quot; at the moment is SDRAM only - the DDR is not planned to be out yet until next year ... so keep that in mind.

I agree about the bandwidth issue, DDR just doesn't have the bandwidth dual-channel RDRAM has. But I would like to know about the timing issues - that is non-sensical.

If anything, DDR has better latency than RDRAM has, so in that respect its better off. However, there shouldn't be any timing issues - the CPU and memory run on two independant (that is, asynchronous) busses - the one hasn't got anything to do with the other.

If you're talking MTH (hello 820 SDRAM boards ) - then I would be very impressed - I doubt that Intel's engineers are that stupid to try that again. MTH as I see it was an attempt at bringing the RDRAM memory-structure into the mobo-market. An RDRAM &quot;adressing&quot; with SDRAM - the memory translator HUB just didn't work - tough .

If you mean something else with &quot;memory timing issues&quot;, then please say so - I would like to know what it is you're talking about .

 

[Jasonh100]

sorry about that

 

[fkloster]

<< If anything, DDR has better latency than RDRAM has.... >>



Interesting Shathal. I believe that any latency penalties inherant with rdram are lost (hidden) when rimms are coupled with i850. I doubt latency is an issue to compare between DDR platforms &amp; i850 platforms. i845? who knows the ramifications of applying a parrallel memory architecture to a platform desinged solely for serial memory. MCH? I don't know...

 

[shathal]

Sorry about what ?!?!?

<Me == confused>

 

[AndyHui]

Latency problems should not be an issue with the Pentium 4 chipsets. Not only was the i850 chipset designed to hide these latencies fairly well, the hardware prefetch also minimises the effect of the very long delays of data going the whole way around the serial RDRAM bus.

Since DDR doesn't have these latencies, then the issue is irrelevant.

As far as I can see with the description of the i845, there is no MTH....support for SDRAM and DDR SDRAM is native, in the same way as it is with the i81x chipsets.

Seriously though, I do not understand why people are complaining about the lower performance with SDRAM and DDR SDRAM as compared to RDRAM.

This situation is exactly the same with nVidia's GeForce 2 GTS and GeForce 2 MX cards. The i845 is &quot;crippled&quot; in exactly the same way the GeForce 2 MX is, as compared to the GeForce 2 GTS and i850 examples. I don't see people complaining about purchasing the GeForce 2 MX.

 

[fkloster]

<< I do not understand why people are complaining... >>



I can answer that. People are complaining because they (...some of us anyways ) don't understand how DDR memory can cripple any platform. The P4 &amp; RDRAM/i850 are a perfect marriage, a communion that NO other (current) cpu/platform/memory subsystem can enjoy. Not that this means a heck of alot though. RDRAM NEEDED to be married off to a platform that is designed for its serial nature...

 

[shathal]

Interesting Shathal. I believe that any latency penalties inherant with rdram are lost (hidden) when rimms are coupled with i850. I doubt latency is an issue to compare between DDR platforms &amp; i850 platforms. i845? who knows the ramifications of applying a parrallel memory architecture to a platform desinged solely for serial memory. MCH? I don't know...

Latency and RDRAM - hidden? (RDRAM would like that) .

RDRAM is &quot;king&quot; in terms of bandwidth, but the RDRAM latency is pretty bad (well - relatively speaking anyway - compared with SDRAM/DDR). SDRAM and DDR, though they have lower Bandwidth (at the moment anyway), have a lot better latency.

Thus, LATENCY-dependant programs run better with DDR. BANDWIDTH-dependant programs/apps run a lot better with RDRAM . RAMBUS would be quite happy if it could get the latency that DDR and SDRAM are getting. And yes, latency does make quite a difference. You can see that, simply by comparing CAS-3 vs. CAS-2 memory .

MCH/MTH shouldn't happen on i845 - at least to the best of my knowledge at the moment.

Want me to look into the matter? .

 

[Marty]

Well, the i845 supports DDR, and Intel is only allowing the support of SDRAM. That tells you that DDR performs sufficiently close to DRDRAM that Intel fears it will cut into the i850's market.

Marty

 

[shathal]

AndyHui making another good post. He's getting a bit of a habit to get there before me. Eville, he is . *grin*

But seriously, Andy's made good points - and Fkloster has answered equally well (thanks for that one, was a &quot;DOH! of course!&quot;-effect for me too) .

DDR is a good thing - it just needs to be used. Despite what AMD say, they are certainly not using it as much as they could (well - if their CPU's were designed FOR DDR). I am quite keen to see the DDR vs. RAMBUS run-ups on the P4-boards, as THAT is the really only platform where latency and bandwidth make BIG impacts.

Hrm - now, where do I get myself a DDR-board from ... doubt that my contact go THAT deep that I can sweet-talk a prototype out of Intel ... shame .

 

[fkloster]

<< Want me to look into the matter? >>




Yes I would like you to present me with current benchmarks that ISOLATE latency in comparison of i850 w/a common DDR platform. I think you will be surprised. Were not talking i820 here you know. i850 latency comparisons with DDR platforms only: ISOLATED...




<< (i845 supports DDR...) ...That tells you that DDR performs sufficiently close to DRDRAM that Intel fears it will cut into the i850's market. >>



Oh boy, thats quite a stretch Marty. I believe Intel is bringing i845 to market SOLELY to offer a lower end/cheeper platform to people.

 

[AndyHui]

shathal: as stated above, the i850 chipset has been designed to compensate for the RAMBUS latencies, along with the hardware prefetch feature on the P4 itself. The latencies are effectively &quot;hidden&quot; from the system and does not affect performance the way it did on the i820 chipset. Latencies are not an issue with RAMBUS on the i850 chipset.

 

[shathal]

i845 will (eventually) support DDR - to start off only SDRAM. Whether this is a marketing ploy to try &amp; hope for RIMMs to be cheap enough to effectively &quot;battle&quot; DDR-sticks at the time, I can't say. (Though I assume it is, at least to a part).

There is one reason I know of, why SDRAM/DDR support only came out AFTER RIMM-support. That is that Intel definately does NOT want a replay of the MTH-issue - they want to get it right this time round, and that takes time. However, it is quite possible that there is some deliberate delaying in the DDR-version of i845 - that is quite possible.

After all, they are thinking like a business...

 

[Noriaki]

i845 aside for a moment...why does everyone think Rambus is the only memory type for the P4?

Lets pretend for the moment nVidia made a Crush 12 for the P4.
That's dual channel DDR, why wouldn't that be able to perform up to par with dual channel Rambus?

This is probably the 10th time I've asked this question and no one ever seems to answer it.

I acknowledge the i845 even with DDR2100 will take a performance hit compared to dual channel PC800, because it's only got 2/3s of the bandwidth.

But I keep getting told (primarily by Fkloster) that dual channel DDR wouldn't perform well with the P4 either. Why?

Assume latency difference don't matter becuase the i850 is designed to hide Rambus' latency (which I believe it does very well).

Bandwidth difference:
Dual Channel PC800 = 3200MB/s, I believe Rambus is roughly 85% bandwidth effecient = 2720MB/s

Dual Channel PC2100 = 4200MB/s, DDR is roughly 65% bandwidth effecient = 2730MB/s

So you get some whopping 10MB/s effective bandwidth difference.
Wowee....so why would the P4 perform any worse with Dual channel DDR than Dual channel Rambus?

 

[AndyHui]

I think that dual channel DDR would be a very good solution. As for economy, I'm not sure that implementing dual DDR is that cheap for motherboard manufacturers.

In terms of performance, apart from nVidia's solution, we haven't seen a dual DDR implementation, so we have no real idea about its real world performance.

 

[shathal]

Yes I would like you to present me with current benchmarks that ISOLATE latency in comparison of i850 w/a common DDR platform. I think you will be surprised. Were not talking i820 here you know. i850 latency comparisons with DDR platforms only: ISOLATED...

WTF you thinking I've been trying to get ahold of . I'm hunting for this for a while.

Noriaki:
That is a good move - dual-channel DDR would definately be a serious contender for RAMBUS (also in terms of price). The memory bandwidth would also be great - that is something that the P4 runs on after all .

So - the question would be, why doesn't Intel do it (I still think the &quot;decision makers&quot; there would prefer to have RAMBUS, rather than DDR). I'm pretty sure that the Intel-engineers have already thought of that option too ... hmmm - we should maybe make a &quot;feeble&quot; takeover of AMD &amp; INTEL, get those two sorted out once &amp; for all *laughs*

But seriously - it's a good idea but I think I doubt that it'll be done by Intel in the near future - maybe VIA or another chipset manufacturer get the idea &amp; do it - THAT would be something to look foreward to.

Could potentially pull the carpet away under RAMBUS' feet - what with Quad-Data-Rate DIMMS sampling at the end of the year ... interesting times .

 

[fkloster]

Noraiki, I have no doubt in my mind that a Dual DDR platform has bandwidth. My question is if it can be implemented to be utilized 'efficiantly' by the P4, which was 'designed' for i850. On paper it looks good, but then again, so did the xbox.

 

[Noriaki]

Oh I have no idea about price.
Dual Channel DDR would need 128 data pins as compared to Rambus's 32. But then a DDR DIMM and a RIMM have similar total pin counts...so would it actually require a bunch more layers for Dual DDR than Dual Rambus? I don't know...I'm not an engineer...

But DDR would definately be cheaper than Rambus...
$62 for 256MB Would that cover the difference in mobo cost?

I dunno...it's just a random thought...

Fkloster: How was the P4 designed for Rambus anymore than anything else?

The P4 uses a 64bit 100Mhz QDR bus..doesn't really resemble the 32bit 400Mhz DDR bus of dual channel Rambus. The i850 handles the conversion between the two different bus widths/speeds.

The i850 is definately designed for Rambus and designed well. But all the P4 needs is a chipset that gives it lots of bandwidth to work with I think.

If you got a chipset that had dual channel DDR supprt, to interleave two 64bit DDR busses into a single 64bit QDR bus for the P4...I think that it would work nicely....

In the case of i815 vs i820 you are dealing with a Sync'd memory bus versus an ASync'd bus. Since the P3 is designed for a Sync'd memory bus, it loses performance. In the P4 example in either case I'm talking about an Asynchronous memory bus. The i850 chipset handles it now, and my theoretical dual channel DDR chipset would have to as well.

 

[shathal]

What Fkloster (I assume) meant that P4 was designed for RAMBUS is BANDWIDTH. It's &quot;nice&quot; that there's similar numbers involved (i.e: 800 and 400 MHz), but as such, bandwidth must have been the determining factor...

 

[Dug]

<< On paper it looks good, but then again, so did the xbox >>


Hey!, don't bring my friend xbox into this.
As Anand said in his nForce review.


<< Remember that the current specs for the XBOX are well beyond even the most advanced products available in the PC market at this time >>

 

[fkloster]

Perhaps your correct Noraiki. If Intel delivers a platform that manages Dual DDR memory well and marrys it up with P4 parts, they may have a winner. Single Channel DDR w/a P4 doesn't stand a chance...



<< Hey!, don't bring my friend xbox into this... >>



Oh, all right...

 

[shathal]

Single channel-DDR doesn't stand a chance.

Well - I am inclined to agree with that. But (ok - technically not DDR), QDR (Quad-Data-Rate) is coming out &quot;soon&quot; - sampling beginning end of this year AFAIK, so this might also be an interesting one.

RAMBUS' next leap will be 1066 MHz RIMMs - I guess Intel will be pairing that up with 533 MHz FSB (133 x 4) for good measure. BUT - I think that QDR might be a serious contender ... now, to see whether it will be actually as good in real-life, as it is on paper .

Anyone from the memory business about to shed some more light on that?