Hands On : Next Generation PC-1066 and RIMM 4200 RDRAM

[burek]

Hands On : Next Generation PC-1066 and RIMM 4200 RDRAM

Too bad Intel's dropping it

 

[ST4RCUTTER]

It's nice and all but Granite Bay will be the death of RDRAM. Greater bandwidth, lower latency, maybe even cheaper...

 

[Athlon4all]

It's nice and all but Granite Bay will be the death of RDRAM. Greater bandwidth, lower latency, maybe even cheaper...

I just would like to inform you ST4R, Granite bay will not provide more bandwidth than PC1066 on 850e (Granite Bay will only support DDR266, so 133x2x2x64/8=4.2GB/ps, the same bandy that PC1066 on 850e provides. Second, PC1066's latency is vastly improved from PC800 and it is much closer to the latency of DDR266 just to clarify. And finally, there is absolutely no way that boards will be cheaper than 850e boards. The cost of running 2 64-bit paths for the memory is signifcantly greater than running 2 16-bit paths on 850e. The cost of the actual chip will be close to that of 850e (due to it being produced on a .18 micron process, while 850e is still at .25) but still, boards will be more. Just to clarify

 

[Ben50]

I think Granite bay will support pc2700 even if not officially.

 

[jeffrey]

RIMM 4200 RDRAM modules are some darned impressive stuff! I never like Rambus as a company, or for their products. However, now I wish Intel would have thrown some support behind the new RIMM4200 standard. Maybe AMD, Transmetta, or VIA can work on some platforms for their chips that use the tech.

 

[Athlon4all]

However, now I wish Intel would have thrown some support behind the new RIMM4200 standard. Maybe AMD, Transmetta, or VIA can work on some platforms for their chips that use the tech.

Well. SiS is making a Dual Channel RDRAM chipset (SiS 658) so even if Intel never makes a new RDRAm chipset again, we will still have a supporter of the technology for the P4.

 

[Scipionix]

Okay, I still don't get it. Why is everyone so excited about RIMM 4200? It's just a physical rearrangement of PC1066 modules so that you need one or two modules instead of two or four. The new Asus board is the same thing as the P4T533-C except with different trace layouts and memory slots with more pins. Please explain.

 

[Bozo Galora]

because traces internal to the modules are far shorter than mobo traces which is extremely important to serial devices
because two slots can be oriented closer to CPU better than 4

 

[Scipionix]

Resulting in actual memory bandwidth that is a few percentage points closer to 4266 MB/s than the 3400 MB/s (80% of maximum) that current 16 bit RDRAM platforms deliver?

 

[Bozo Galora]

first of all, its about making it work.
DDR-II is flopping big time - many unforseen probs
rambus is way past them on WORKING technology
not to mention RDRIMM 9600 is on its way link

secondly, and I'm at a great disadvantage here because I have no P4T533 board to investigate.......
but to speculate here:
RIMM 1066 has two channels of 2.1GB/s for a total of 4.2GB/sec
RIMM 4200 has two channels EACH of 4.2GB/s for a total of 8.4GB/s
(serial technology)
Note on the P4T533 test setup, they only used 1 512 stick
The question is whether the new ASUS board will allow the 8.4GB bandwidth, and whether I am correct on the 8.4GB/s reasoning.
-----------------------

but that is just my opinion, and I could be wrong

 

[Scipionix]

No, it is still two 16 bit channels, the only difference is that both channels access each 232 pin RIMM. It's the same chipset and the same memory technology, the only difference is that memory for both channels is physically present on each module.

 

[RaynorWolfcastle]

Originally posted by: Scipionix
No, it is still two 16 bit channels, the only difference is that both channels access each 232 pin RIMM. It's the same chipset and the same memory technology, the only difference is that memory for both channels is physically present on each module.

Making it PC4200 more easily upgradeable and more likely to support slots for more RAM

-Ice

 

[robg1701]

RIMM 4200 is just a 32bit rimm, in that it has 2 16bit rdram channels on it, rather than current setups which use 2 seperate 16bit rimms on their own channel....it delivers 4.2GB of bandwidth, hence the name. Having 2 32bit RIMMS ddoesnt double the bandwidth versus one, just like havinf 4 16bit rimms doesnt double the bandwidth versus having two....so the only benefit is not needing to install in pairs.

As for Granite Bay...im sure it will have at least unoffoicial PC2700 support....though we know its consumer 'springdale'version defiantely will since its intended to support the 667FSB Prescott at launch...

 

[Athlon4all]

Well, about why everyone is excited. I'm not as excited as I had been, but still it rocks.

As for Granite Bay...im sure it will have at least unoffoicial PC2700 support....though we know its consumer 'springdale'version defiantely will since its intended to support the 667FSB Prescott at launch...

I disagree. Yes Springdale will have DDR333 support, but you have to understand the amount of validation that Granite Bay would need to go under in order for Intel to safely even unoffically support DDR333. Plus, there is absolutely no point because the fsb will still be at 533fsb.

 

[robg1701]

Well, the i845G can do it, and the FSB is still only a 533FSB.....so why not a new chipset 6 months down the line ?

 

[Pabster]

Bozo Galora wrote:

"first of all, its about making it work.
DDR-II is flopping big time - many unforseen probs
rambus is way past them on WORKING technology
not to mention RDRIMM 9600 is on its way"

Where are you hearing DDR-II is "flopping big time"? Last I saw, Samsung already has 900MHz+ DDR-II modules in production and being validated. And rumor has it 1GHz parts are already being sampled as well.

"secondly, and I'm at a great disadvantage here because I have no P4T533 board to investigate.......
but to speculate here:
RIMM 1066 has two channels of 2.1GB/s for a total of 4.2GB/sec
RIMM 4200 has two channels EACH of 4.2GB/s for a total of 8.4GB/s
(serial technology)
Note on the P4T533 test setup, they only used 1 512 stick
The question is whether the new ASUS board will allow the 8.4GB bandwidth, and whether I am correct on the 8.4GB/s reasoning"

A common misconception. The 32-bit RIMM (as others have pointed out, but I'll reiterate) simply banks the RAM on one physical module. There is not a "doubling" of bandwith from the legacy 16-bit; RIMM 4200 is still 4.2GB/s. I still firmly believe DDR-II will overtake RDRAM, and Intel seems to be convinced of it as well. (Not that they haven't been wrong before, but...)

 

[ST4RCUTTER]

I just would like to inform you ST4R, Granite bay will not provide more bandwidth than PC1066 on 850e (Granite Bay will only support DDR266, so 133x2x2x64/8=4.2GB/ps, the same bandy that PC1066 on 850e provides.

Nothing is stopping people from using PC2700 and an overclocked FSB. I would also be surprised if GB or a dual-channel chipset from another maker does not support DDR333.


Second, PC1066's latency is vastly improved from PC800 and it is much closer to the latency of DDR266 just to clarify.

With how many banks filled? Regardless of how fast the memory is ramped, the serial nature of RDRAM will always include wait states to keep all the RIMMS in step. I would like to know how low the memory makers/board makers have managed to reduce latency between RIMMS though. In stating reduced latency comparable to DDR, what exactly are they comparing?


And finally, there is absolutely no way that boards will be cheaper than 850e boards. The cost of running 2 64-bit paths for the memory is signifcantly greater than running 2 16-bit paths on 850e. The cost of the actual chip will be close to that of 850e (due to it being produced on a .18 micron process, while 850e is still at .25) but still, boards will be more. Just to clarify [/i]

No arguments there. Unless VIA has an option in volume that can offer a similar product at less cost.

 

[Athlon4all]

Well, the i845G can do it, and the FSB is still only a 533FSB.....so why not a new chipset 6 months down the line ?

Here's why. One a Dual Channel DDR Memory controller is much more complex than a Single CHannel DDR controller, second, 845G its different. DDR266 can hardly saturate the 533fsb, and DDR333 still can't really. With Granite Bay however, with DDR266, you will have 4.2GB/ps, enough to saturate the fsb. You see the difference?

Where are you hearing DDR-II is "flopping big time"? Last I saw, Samsung already has 900MHz+ DDR-II modules in production and being validated. And rumor has it 1GHz parts are already being sampled as well.

I have to agree. I haven't heard anything of the sort.

A common misconception. The 32-bit RIMM (as others have pointed out, but I'll reiterate) simply banks the RAM on one physical module. There is not a "doubling" of bandwith from the legacy 16-bit; RIMM 4200 is still 4.2GB/s. I still firmly believe DDR-II will overtake RDRAM, and Intel seems to be convinced of it as well. (Not that they haven't been wrong before, but...)

Again, while I am not sure about DDR-II overtaking RDRAM, this is true (which I didn';t realize at first either). The 32-bit RIMM still requires support from the chipset. Basically, the RIMM 4200 is split in half and in the mobo traces layout, one of the channels goes to one side of the RIMM slot while the other channel goes to the other. This is as you can see why RIMM4200's don't need to be installed in pairs.

EDIT:

With how many banks filled? Regardless of how fast the memory is ramped, the serial nature of RDRAM will always include wait states to keep all the RIMMS in step. I would like to know how low the memory makers/board makers have managed to reduce latency between RIMMS though. In stating reduced latency comparable to DDR, what exactly are they comparing?

I honestly don't know. But to be quite honest, I have always really thought that the penalty of having additional RIMM's never resulted in as big of a performance drop as people think. Just my opinion.

Nothing is stopping people from using PC2700 and an overclocked FSB. I would also be surprised if GB or a dual-channel chipset from another maker does not support DDR333.

Agreed on another manufaturer making a DDR333 supported DC DDR chipset, but as for ocing, well, yes but that isn't always viable, and Granite Bay may or may not be able to handle it.

 

[robg1701]

Originally posted by: Athlon4all

Well, the i845G can do it, and the FSB is still only a 533FSB.....so why not a new chipset 6 months down the line ?

Here's why. One a Dual Channel DDR Memory controller is much more complex than a Single CHannel DDR controller, second, 845G its different. DDR266 can hardly saturate the 533fsb, and DDR333 still can't really. With Granite Bay however, with DDR266, you will have 4.2GB/ps, enough to saturate the fsb. You see the difference?

Nope....KT400 certainly isnt gonna make a huge difference over KT333, which was hardly an improvement over KT266A...the Athlon is finding the option of higher mem speed than FSB speed, so why not the P4...just cause it isnt required doesnt mean it wont happen, and since intel will be releasing officially supported PC2700 versions of todays E and G chipsets after summer, again why shouldnt this apply (officially or not) to Granite Bay?

Also, if dual channel DDR is so hard, then how come nVidia has had it out so long now on a 4-layer PCB ? This also relates to an earlier post where you said it would be far more expensive than normal and thus Granite Bay boards wouldnt be cheaper than RDRAM boards....based on current trends i cant say i believe that in the slightest either, since in the UK at least, the cheapest i850(e) board i can find is certainly nowhere near as cheap as the DDR based boards, or the dual channel nForce either (granted at introduction the nForce cost too much, but a huge chunk of that was the chipset itself...theyve massively come into line with the competition since).

 

[Athlon4all]

Nope....KT400 certainly isnt gonna make a huge difference over KT333, which was hardly an improvement over KT266A...the Athlon is finding the option of higher mem speed than FSB speed, so why not the P4...just cause it isnt required doesnt mean it wont happen, and since intel will be releasing officially supported PC2700 versions of todays E and G chipsets after summer, again why shouldnt this apply (officially or not) to Granite Bay?

Its a little bit different with the P4. It's fsb is not being pushed to the max by DDR at its current levels, meanwhile, DDR266 fulled saturates the Athlon's fsb. And further, Intel is very in depth when it comes to validating their products and they may want to hype the Springdale launch by giving DDR333 as a big feature added to an allready impressive product. We shall see.

Also, if dual channel DDR is so hard, then how come nVidia has had it out so long now on a 4-layer PCB ? This also relates to an earlier post where you said it would be far more expensive than normal and thus Granite Bay boards wouldnt be cheaper than RDRAM boards....based on current trends i cant say i believe that in the slightest either, since in the UK at least, the cheapest i850(e) board i can find is certainly nowhere near as cheap as the DDR based boards, or the dual channel nForce either (granted at introduction the nForce cost too much, but a huge chunk of that was the chipset itself...theyve massively come into line with the competition since).

I really cannot say. To be honest, There is a lot of confusion around on whether nForce is even a true Dual Channel DDR chipset. And further, something you must keep in mind is that nForce is being produced on a .15 Micron process meanwhile Granite Bay will be on a .18. Now I do agree that this doesn't change (as I've already pointed out) the costs of producing a Granite Bay board, but it is worth mentioning. Oh and about 850e boards prices, well, I will admit that this isn't an 850e board, but Abit's TH7II (which supports 533fsb and PC1066 unoffically) is avialable here in the US for $119, when for example, Abit's BG7 costs $130.

 

[robg1701]

Its a little bit different with the P4. It's fsb is not being pushed to the max by DDR at its current levels, meanwhile, DDR266 fulled saturates the Athlon's fsb. And further, Intel is very in depth when it comes to validating their products and they may want to hype the Springdale launch by giving DDR333 as a big feature added to an allready impressive product. We shall see.

What makes it any different ? DDR333 more than saturates the Athlons bus yet here they come adding more in DDR400 support - when its well known the Athlons not going to get a faster bus for a long time to come. You are correct that theyll hype it for everything its worth with springdale, but i cant see them not having it in there 6months from now when theyre ready to launch al their official DDR333 chipsets at last, with some current products already supporting it. For example, the only reason i can see i845G having DDR333 support when i845E does not, is that the E is just a direct respin of the i845D whereas i845G involved a new GMCH. If the same happens with Granite Bay -> Springdale, it will be in there to begin with. As i said, i feel it will at least unnofficially support it.

I really cannot say. To be honest, There is a lot of confusion around on whether nForce is even a true Dual Channel DDR chipset. And further, something you must keep in mind is that nForce is being produced on a .15 Micron process meanwhile Granite Bay will be on a .18. Now I do agree that this doesn't change (as I've already pointed out) the costs of producing a Granite Bay board, but it is worth mentioning. Oh and about 850e boards prices, well, I will admit that this isn't an 850e board, but Abit's TH7II (which supports 533fsb and PC1066 unoffically) is avialable here in the US for $119, when for example, Abit's BG7 costs $130.

Confusion about nForce being true dual channel ? Either way i wouldnt say it matters, theres 2 DDR channels in effect (the effect on the integrated video between one and two channes is clearly visible), so the routing issues are there at the very least.

Is that the TH7-II or the TH7-II RAID ? Since it wouldnt be very fair to compare the rather feature packed BG7 (onboard raid, sound, video, lan if i recall..a fair increase in material costs) to just the plain old TH7-II (Sound ? Nothing ? cant remember), especially when theyre not really aimed at the same market area IMO. Also, whereas the TH7 II is now quite an old product, the BG7 is a new board so id expect a little scope for price reduction.

As for 533/PC1066 support, that requires a bit of luck with the DCRGs since they are not all the newer ICS chips if im not mistaken, meaning that in making it to PC1066 speeds your not going much higher without dropping the rdram multiplier or modifying the board. In addition, unless a bios change fixes it, many people have been saying theyre finding that the only way to make the Samsung PC1066 work in the TH7-II is to stick some PC800 in along with it, sort of defeating the point.

 

[MadRat]

<<...many people have been saying theyre finding that the only way to make the Samsung PC1066 work in the TH7-II is to stick some PC800 in along with it, sort of defeating the point.>>

Please clarify the source.

 

[Degenerate]

Perhaps VIA will make support...But somehow i find that hard.

 

[robg1701]

Originally posted by: MadRat
<<...many people have been saying theyre finding that the only way to make the Samsung PC1066 work in the TH7-II is to stick some PC800 in along with it, sort of defeating the point.>>

Please clarify the source.

In the forum for the UK company OcUK (Overclockers UK) http://www.overclockers.co.uk its been talked about a few times by a few different people, some from the company, some not, that the only way to get the PC1066 working properly/at all was to pair it with some PC800.....some of the guys are very experienced overclockers, and while i wouldnt trust just one of em, when several of em find the same thing it suggests a pattern. I should have mentioned that yesteday i glanced at somthing where someone suggested that Kingston PC1066 worked in the TH7-II, though i cant find it...and unfortunately the search option isnt available.