XDR vs DDR

chaphack

Veteran
http://www.reed-electronics.com/ele...icleId=CA310194&text=rambus&rid=0&rme=0&cfd=1

XDR vs. DDR


By Alex Romanelli -- Electronic News, 7/10/2003


Rambus Inc.’s strategy for its next-generation XDR DRAM technology relies on its ability to eclipse DDR DRAM in the marketplace. The Los Altos, Calif.-based IP company has designed XDR (formerly code-named Yellowstone) from the ground up, claiming to address the shortcomings of the DDR architecture.
Rich Warmke, marketing director of Rambus’ memory interface division, sees Moore’s Law as driving on-chip integration and speeds, but interface technology is not keeping pace.


“Pin counts are continuing to go up but at a much slower pace than Moore’s Law,â€￾ he said. “As you get more and more transistors on single dies, while I/O counts are not going up quite as fast, the I/O-to-gate count ratio is plummeting. The end result is that per pin bandwidth is skyrocketing. Logic is pushing into multi-gigahertz signaling rates. So you need to go to differential signaling since single-ended signaling is not as immune to noise. Topologies also start top be more limited as you get up to higher speeds, all the high-speed logical interface links are point-to-point. We see that applying to DRAM. So XDR is going to be the first high-volume DRAM that has a differential data point-to-point interface.â€￾

Warmke said PC main memory is on a slower trajectory but it is also following this trend, so over time PC main memory will catch up as increasing bandwidth is required between the controller and the frame buffer. Rambus estimates that by 2005/2006, the PC will require bandwidth of 50Gbyte/sec. and by 2010 maybe 500Gbyte/sec. Customers are already demanding multi-gigahertz speeds beyond the capabilities of today’s DDR and RDAM, he said.

“There is no other DRAM that can get close to this level of performance, so we’re pretty far ahead of our competition, and we’re seeing customers wanting this bandwidth level,â€￾ Warmke said.

XDR is effectively a hybrid of DDR and Rambus DRAM, designed to combine the best elements of both.

“Over time we think that PC main memory systems are going to start requiring more and more bandwidth beyond what our DRAM is able to provide at the high-end today,â€￾ he added. “This topology has never been done before, where you have a point-to-point differential data bus and a bussed single-ended address bus. It affords the best of both the parallel approach used with DDR and the bus approach used with RDRAM.â€￾

XDR’s point-to-point signaling in the data bus set it apart from other memory technologies and allow it to reach speeds up to 10GHz.

XDR or DDR?

Part of XDR’s success depends on the failure of DDR technology to keep pace with industry demands. DDR and DDR2 have widespread industry support right now, from the likes of Intel Corp., but Warmke sees a time when DDR’s single-ended signaling just won’t cut it anymore.

“We do not see single ended signaling being a reliable or robust means of getting bits on or off chips much beyond 1.6GHz, and even that is a big challenge,â€￾ he said. “Even the GDDR family is going to have to go to differential signaling. With XDR we are clearly out ahead.â€￾

Whether or not Rambus is out ahead will be key to its success in infiltrating the main memory of mainstream PCs.

“XDR is very good technology but it has some key limitations -- basically expandability -- that will keep some people from wanting to bother with it,â€￾ said Peter Glaskowsky, editor in chief of In-Stat/MDR’s Microprocessor Report.( :LOL: ) “ If three years from now it’s still obvious that DDR is still meeting the needs of the industry, then very few people will choose to make a huge strategic shift to XDR. And I think that is likely to be the case.â€￾

Just as Rambus’ XDR strategy resembles its strategy for RDRAM in the late 1990s, so does XDR’s technical advantage.

“XDR does have the same advantage that RDRAM had several years ago -- highest bandwidth per pin,â€￾ said Steve Cullen, an analyst with In-Stat/MDR (which is owned by Reed Business Information, the parent company of Electronic News.) “A few years ago, the number of DRAMs in a PC would decrease because the average memory density size was not growing as fast as DRAM densities were, and you had this granularity problem. You needed more pins but the only way to get more pins was to put in more memory than you really needed.

“If memory requirements for PCs didn’t keep up with DRAM technology, you would have this granularity problem, and that’s when Rambus makes sense. That’s why they’ve got success in consumer applications, they have the granularity problems because they only have one or two DRAMs.

“If the granularity problem comes back, if PC memory sizes stop growing at the same rate that DRAM densities are increasing, then at some point the Rambus concept would make sense. I can say with absolute certainty that’s not going to happen in the next three to four years, which means its unlikely to happen in five. But if you go out seven or eight years, it’s conceivable.â€￾

Building Blocks

XDR consists of three physical layer building blocks: Rambus’ FlexPhase, a controller based circuit technology; differential Rambus signaling levels (DRSL): and octal data rate signaling.

“For every pair of data to be routed at different lengths to the DRAM, FlexPhase allows a new level of freedom to the DRAM designer. As you get into multi-GHZ signaling rates you can’t guarantee that your PCB trace lengths or timing error in the system are going to be fixed. FlexPhase allows for a very precise centering of data with on-chip clock without having to trace length match signals.â€￾

DRSL is a very high-speed low power signaling interface, targeted at standard four layer PCBs at 50 ohms impedence. It’s a bi-directional differential technology. Octal data rate signaling allows 8 bits to be transmitted per clock cycle, which allows XDR to achieve a 256Mbit device at 3.2GHz data rates with a 400MHz clock. Warmke said the rest of the memory industry is still operating with single rate and double data rates. Rambus claims XDR is scalable to allow, in time, an 8Gbit device running at 6.4GHz.

Rambus is banking on its first mover advantage with these new memory architecture techniques.

“If the need for bandwidth continues to increase there are not going to be a whole lot of alternatives out there that do not go toward differential signaling and the point-to-point topology,â€￾ Warmke said. “Given our head start, we’re going to be an extremely compelling solution to match that need.â€￾


Seems like 52GB/s is ~around~ there for 2005/6. Anyone knows how GDR3 be faring? Heard its ~around~ there too. Guys like ATi are favoring GDDR3 yeay?
 
This topic most likely belongs in one of the other forums.

But I don't think you'll see XDR come to the PC market, more just closed systems such as gaming consoles or something of that nature.

It has the performance edge, but your not going to see it inside your PC probably ever. Rambus I think is done when it comes to the PC market. Best stick to closed boxes.
 
I *think* it can give us a rough gauge of comparative bandwidth or something, provided Rambus estimate is right. PS2 was at 3.2GB/s at a time when PC are still in the 1~2GB/s range and Xbox was at 6.4GB/s until dual channel DDR came along to the PC a year or 2 later.

Of coz, thats what i *thinketh* only, any 3D hardware dude care to say something?
 
I dunno...maybe a rough gauge of future console bandwith and work from there how the 3D be liketh? i think...:?: :LOL: :?:
 
If RAMBUS were smart, their PC strategy should focus on making XDR follow DDR route - first provide their RAM for video cards, where bandwidth requirements far outstrip any other aspect of PC architecture and once XDR received widespread adaptation and acceptance there, migrate it to system RAM market.
 
PS3 XDR DRAM will be atleast 26GB/s(as this is XDR's lowest capability). Who knows, maybe if PS3 launches later they could put in the 50GB/s DRAM versions.

XBOX2 I don't know what type of memory, DDR2? But most likely 40GB/s.

Then again you will have e-DRAM(Internal memory) for ps3 which could hit easily 100+GB/s.. So comparing Xbox2 and PS3 main DRAM won't get you far.
 
The article pretty much can be summarized as this:

"XDR is technically superior, in terms of bandwidth per pin. Bandwidth per pin is becoming increasingly crucial as pin counts are not keeping up with transistor counts. The main obstacles to acceptance of XDR are: 1. Lack of expandability 2. Lack of industry support. Rambus is banking on the failure of DDR to meet future requirements, and by being a first-mover, it hopes to make XDR the dominant next-gen memory format."

Both of this obstacles are not an issue for PS3.

Expandability - PS3 will have standardized hardware, no need for expandability.

Lack of industry support - If history is any guide, PS3 will sell in the tens of millions. It can acheive economies of scale by itself. Even if no one else uses XDR, the price of XDR will be very close to the marginal cost of production, just by the sheer volume PS3 provides.

IMHO, the greatest risk of XDR is that it is not (to my knowledge), a proven technology. Sony needs to make sure XDR can deliver the performance they need.
 
PS3 main memory bandwidth guesstimate:

lowend: 25~26 GB/sec

midrange: 50~52 GB/sec

highend: 102~104 GB/sec



I'm hoping for midrange and 512 MB.


lowend 25~26 GB/sec and 256 MB seems a bit weak.
 
I think another reason why XDR won't at least be immediately favoured over DDR2 based solutions in the PC market stems from CPU design. As of now, lower latency seems more important for archieving higher real world performance than 10-20% of increased bandwidth for most applications (thanks to ever larger caches). OTOH with memory controllers being integrated into CPUs, per pin throughput might become more important, thus favouring rambus a bit. Yet chances are that MRAM will have matured in the 2010 timeframe and made obsolete capacitor based memory.
 
PiNkY said:
I think another reason why XDR won't at least be immediately favoured over DDR2 based solutions in the PC market stems from CPU design. As of now, lower latency seems more important for archieving higher real world performance than 10-20% of increased bandwidth for most applications (thanks to ever larger caches). OTOH with memory controllers being integrated into CPUs, per pin throughput might become more important, thus favouring rambus a bit. Yet chances are that MRAM will have matured in the 2010 timeframe and made obsolete capacitor based memory.

To be fair XDR should have considerable lower latency compared to Direct RDRAM.
 
XDR is unlikely to be all that superior in bandwith per pin in the end, I have been pretty amazed at what DRAM manufacturers have been able to do with GDDR2 when pushed ... yellowstone's advantage is powerconsumption.

Of course they are both utterly useless for PCs. I dont buy the "we wont need extendability somewhere in the future" arguement ... when that comes true we will simply use eDRAM or some type of MCMs instead, and for the latter yellowstone is far from optimal.
 
Megadrive1988 said:
lowend 25~26 GB/sec and 256 MB seems a bit weak.

Interestingly enough this was actually the rumor that appeared this summer. It's also about where the high end video card segment is now, in 2003. By comparision, five years ago in 1998 high end video cards (nVidia TNT) had only 16MB of memory and 1.8GB/sec of bandwidth (same point in PS2's lifespan as it is now in PS3's). If this is the PS3 external memory config, then they really are in some trouble...
 
Well to be fair, while 26GB/s might sound slow come 2005; you also have to remember that PS3 will have loads of e-DRAM on it's CPU. Unlike PS2.
 
PiNkY said:
Yet chances are that MRAM will have matured in the 2010 timeframe and made obsolete capacitor based memory.

Don't count on it. This branch is still in its early stages...
 
Paul while bandwidth considerations are rather moot as long as the whole cahing system works out. by 2006 I imagine (the way things are going) top end GFX boards are gonna be carrying something approaching 1GB of memory alone.

256 really isn't enough. unless Sony really want Devs to start using some rather extreme memory management :oops:
 
Dude this is why PS3 will have a ton of e-DRAM on it.

Your talking a good 16-32MB of it on the CPU, and more on the GPU.

256MB is fine, concidering that they plan to shove alot of e-DRAM on PS3. If there was no e-DRAM, there would be a problem.
 
Dude this is why PS3 will have a ton of e-DRAM on it.

like IO said, this will be an good way to maxmise bandwidth usage (that is what we are talking about in regards to eDRAM right?). however storage for object/media/texture/etc data isn;t help by this in the slightest.

Your talking a good 16-32MB of it on the CPU, and more on the GPU.

256MB is fine, concidering that they plan to shove alot of e-DRAM on PS3. If there was no e-DRAM, there would be a problem.

that eDRAM (if I'm looking at this correctly) will act as a L3 stractch space for data duplication, unless you plan on using some sort or swapping system when fetching from sys ram it's not gonna caompare well (capacity wise) against GFX cards in that period.

note: I'm ignoring tecture compression (it's rather moot) and procedural texturing (useful for aonly a subset of problems) for the sake of simplicity here.
 
Remember what the classic Playstation architecture is though, smaller ammounts of faster DRAM.

In the end though e-DRAM is DRAM. And it will be used as such.

While it is true that you won't see huge ammounts of 3d game data(vertices, texture) stored in this e-DRAM; with the huge bandwidth though, why keep it in for long?

This is the classic battle between larger or faster, and we'll just have to wait and see what happenes.
 
Back
Top