Intel 845 DDR Chipset Review

[Wombat]

But that's one of RDRAM's greatest advantages! It's easy *and* cheap to make a dual-channel configuration.

Cheap, yes. Easy? Hell no. Trying to get the timing right for RDRAM is a nightmare. Those serial lines run so fast it's sometimes amazing RDRAM works at all. Yes, there are fewer traces, that's why RDRAM is good in things like consoles. However, the latency absolutely sucks, and there's nothing that can be done about that - it's a part of the RDRAM design.

Dual-channel DDR is harder to design/manufacture and will hence be more expensive.

The short answer is ummmm...No. The RAM itself wouldn't be dual-channel. It wouldn't care. It's the motherboard that cares. More traces, some more logic, yeah. But not the RAM chips. As far as being harder, maybe. But I can think of at least 3 chipsets that do it. One of them is even consumer market, it's made by nVidia of all people.

[isochar]

You know I was talking about motherboard design. The DDR dimms are easier to manufacture than RIMMs, but the opposite holds true for mobos - especially dual-channel.

However, the latency absolutely sucks, and there's nothing that can be done about that - it's a part of the RDRAM design.

Have you even taken a look at Ace's review of RDRAM at PC1066? It has lower latency than DDR or SDRAM! Throw those RIMMs into a dual-channel configuration and it'll smoke DDR 2700.

The short answer is ummmm...No. The RAM itself wouldn't be dual-channel. It wouldn't care. It's the motherboard that cares. More traces, some more logic, yeah. But not the RAM chips. As far as being harder, maybe. But I can think of at least 3 chipsets that do it. One of them is even consumer market, it's made by nVidia of all people.

Ah, and my point is proven. The problem with DDR is that it's signal integrity significantly degrades when it's memory is fully populated. That's what the huge controversy with the nvidia boards is all about, to no fault of nvidia.

The only inherent problem I can think of with RIMMs is the latency, but that will be solved in the upcoming months...

[Paulr]

Well exactly.
The next generation of RAMBUS will feature lower latency and a manufacturing process that greatly reduces the price.
RIMM based systems are still going to be the better performers when it comes to P4's with the DDR option replacing SDRAM as the entry point configuration.

[Novdid]

Dual channel pc2100 will perform better on a P4 than dual channel pc1066.

[Wombat]

You know I was talking about motherboard design

When you said "easy," that pretty much ruled out RAMBUS MB design.

Ah, and my point is proven. The problem with DDR is that it's signal integrity significantly degrades when it's memory is fully populated. That's what the huge controversy with the nvidia boards is all about, to no fault of nvidia.

It IS nVidia's fault. I've seen much, much, much more DDR in a system chugging along perfectly happy.

The only inherent problem I can think of with RIMMs is the latency, but that will be solved in the upcoming months...

No it won't, and no it isn't. RDRAM is hindered by the way the RAM chips are put in series. Some data has to pass through the entire RDRAM data path before it even leaves the RIMM. Can they make the chips faster? Sure, but so can the SDRAM crowd, and they have the advantage of running chips in parallel, rather than in series.

And the real bitch of working with RDRAM is still the timing. Keeping a 400 or 500MHz signal intact and sensible across a MB sucks ass, big time. I'd much rather work with SDRAM.

[isochar]

No it won't, and no it isn't. RDRAM is hindered by the way the RAM chips are put in series. Some data has to pass through the entire RDRAM data path before it even leaves the RIMM. Can they make the chips faster? Sure, but so can the SDRAM crowd, and they have the advantage of running chips in parallel, rather than in series.

http://www.aceshardware.com/read.jsp?id=45000217

This page specifically:

http://www.aceshardware.com/Spades/read.php?article_id=45000219

[Wombat]

Okay, I read it. What's your point? The only thing that even beat 133MHz SDR DRAM for latency was PC1066 RDRAM. You're comparing tech that hit the streets years ago to tech that isn't even out yet. That chart is meaningless as far as actually comparing the two technologies.

[Technoid]

Just wwant to point out one thing....

The rising prices of DDR ram is because of Christmas and nothing else....

It will sink when the holidays are over...

[Novdid]

A dual channel DDR interface will have latencies better than dual rimms running at 1066MHz, the nature of RDRAM cannot be changed, maybe tweaked to the better though. The single channel pc2700 interface on the sis 645 sometimes manage to outpace the i850, and that with having significantly less bandwidth. Imagine what the P4X266A would do to i850 if the memory were running at 166MHz. I have to say that VIA although they always manage to design lousy memory controllers really managed to design THE best DDR controller (KT266A, P4X266A) that is available to the public.

RDRAM is to expensive and will always be to expensive, it's to innefficient, and it's to complicated to design a good memory controller it seems. The idea of having RDRAM on a graphic board is ridiculous.

The only future I see for RDRAM is in consoles, where DDR also is gaining ground.

My very uneducated conclusion is that RDRAM will die a long and painful death as Intel and Nintendo are showing Rambus their backs.

[Wombat]

RDRAM might still have some place. It does have the low trace count going for it, but that's about it. The weak-signal and differential signalling are nice technologies, just used in the wrong way.

So, single chips are good for handhelds and consoles. For certain computing tasks it might be worthwhile, if they improve their timing, and when (if not already) they come out with lower-heat chips, and can repair RIMMs. My computer is going to stay with SDRAM for a while though.

[superfly]

There is another developement that should come relatively soon as far as rambus memory tech goes,they already have evaluation samples of rambus memory using a 32 bit interface and the memory itself is running at 533 mhz....

Now correct me here if i'm wrong,but if intel chooses to go with that solution and uses a dual channel memory controler just like the existing 850 does on their following chipset,that would not only cut latency issues in half,but also provide about 8.4 gig/sec worth of bandwith,which is much faster that any dual channel sdram ddr could be.


You'd need the sdram ddr modules themselves run at least at 250 mhz(500 mhz DDR) in a dual channel 64 bit architecture to match the performance of a dual channel 32 bit rambus memory running at 533 mhz(pc 1066).

[rylan]

Let me add some info, since I design boards at work and have to deal with memory interfaces.

Rambus has extremely high internal bandwidth (higher than 266MHz DDR). However, the databus width to get in/out of the memory is a pathetic 8 (or 16) bits wide. So with dual channel at best you're looking at a 32bit wide memory bus.
As an example, thats like trying to get the traffic from an 8 lane super highway to all exit off of a 1 lane ramp at rush hour.
Rambus excels at streaming data (video editing and capture) and blows goats for random acccess (most computing, games, etc).

Rambus interface design blows chunks.. we've looked possibly using rambus for soemthing like a high speed framebuffer, but the board layout and termination is just wacky. For motherboards, unused slots have to be terminated (the lovely stick thats in the 'empty' slots).. plus the speed of rambus is load-dependant.. meaning that as you add more chips, the overall speed actually DROPS.. this is partly due to it using serialized data, where you have to wait for the data to get to the sockets with the longer traces... not to mention the EMI shielding issues (gotta love the metal slab over the memory modules).

I've got a lot more technical info if ya want.. feel free to ask.
The bottom line is though, rambus is a joke to 99% of the people in the design industry.. even the sales reps for distributors laugh when we ask if anyone is buying rambus. There are a lot of cool DRAM technologies that are sampling parts now (QDR, 400MHz SDRam etc). There is so much negiatve sentiment towards rambus in the industry, and other superior memory designs, that I don't expect to see it around for very long.

[Indiana]

Originally posted by superfly
There is another developement that should come relatively soon as far as rambus memory tech goes,they already have evaluation samples of rambus memory using a 32 bit interface and the memory itself is running at 533 mhz....

Yes, sure. And Transmeta have their TM5800 CPU available, and BitBoys also have samples of their Avalanche chip, and Amiga has their AmigaOne out...
Sorry, but this is just plain rumours/vapourware until someone outside actually saw such a thing working. And even if, there's still the question if Rambus has the financial means to bring it to the market.

[superfly]

Actually there was....


I saw an article over at ace's hardware some months ago where pictures were shown of the memory itself,it didin't have a single large heat spreader anymore,it was using individual heat sinks for the memory modules themselves and the number of traces on the pcb itself was also much higher than what we see in the current version.


Again they don't make the chips,they just licence the technology to companies who want to produce it.


And besides,regardless of it's disadvantages,it's still one the fastest types or memory around for the p4 and only now,over a year since the 850 has been introduced,do we see DDR chipsets getting close to it's performance in real world apps.


Once pc 1066 if officially introduced in a few short months,it'll regain the lead again,no if's and but's about it and no one can claim that RDram is way more expensive than DDR sdram because it's no longer true.

I'm sure most saw the anandtech article a couple of months ago where they tested 533 mhz RDram and on modified asus p4t,which only needed different clock generators to support the memory,and overall,the performance increased about 15~20% over standard pc 800,and that was on a willamette p4,who knows how much better it might perform on the 0.13 micron version due out in a couple of weeks....


For instance in my area,256 megs of RDram is about 30$ more than the same amount of DDR sdram pc 2100,so price is pretty much no longer an issue anymore.

[Nuno]

But that's one of RDRAM's greatest advantages! It's easy *and* cheap to make a dual-channel configuration. Dual-channel DDR is harder to design/manufacture and will hence be more expensive.

That´s correct. A dual-channel RDRAM is nothing more than a 32-bit bus They could actually do a quad-channel RDRAM and that would equal the traces needed for the 64 bit bus SDRAM uses. The trick on RDRAM is the insane operating speeds, but I always found a 16-bit bus too much of a handicap.

Apart that and the price, there´s nothing intrinsecally wrong about RDRAM. SDRAM still has a lot of space to grow, next step will be DDR-333 and later QDR.