Rambus XDR also on Graphic Cards?

Farid

Artist formely known as Vysez
Veteran
Supporter
Link to X-bit Labs

Rambus may come back onto the market of mainstream hardware for personal computers with its XDR DRAM memory technology on the forthcoming generations of graphics cards. Sources close to the company said there have been negotiations between graphics chip companies and Rambus amid some memory makers’ hopes on the ramp of XDR for graphics applications.

XDR for Next Breeds of Graphics Processors?

XDR memory’s novel system topology allows point-to-point differential data interconnects to scale to multi-gigahertz speeds, while the bussed address and command signals allow a scalable range of memory system capacity supporting from 1 to 36 DRAM devices.

XDR DRAMs will be available in multiple speed bins, device densities, and device widths. With densities ranging from 256Mb to 8Gb, and device widths ranging from x1 to x32, XDR DRAM satisfies the needs of both high-bandwidth and high-capacity systems, the company indicated.

Toshiba has been sampling 512Mb XDR DRAM 3.20GHz devices since late 2003 and is on track to deliver the actual products in 2005 or 2006. XDR roadmaps currently include devices with up to 6.40GHz core-clock.

Sony’s PlayStation 3 is expected to use XDR DRAM and is currently the most well-known XDR-based product.

XDR Inevitable for Graphics, Consumer Applications?

“Graphics seems to be one of the important initial targets for XDR, as graphics applications today have nearly unlimited need for bandwidth out of a single DRAM. Networking is another important market as networking cards need high bandwidth but low capacity,†a source close to Rambus and GPU makers told X-bit labs.

Representatives for Toshiba recently said that small amount of XDR memory for graphics applications will be available already next year. In 2006 equal volumes of XDR will be sold for consumer applications, including Sony PlayStation 3, and graphics devices. According to Toshiba, networking devices will use XDR memory in 2006, but it is unlikely that XDR will be in extremely high demand from networking companies: graphics and consumer electronics XDR markets will be much larger.

Since the data that is stored in a networking card packet buffer does not stay around long – it is sent out relatively quickly – there is no need for high capacity. High bandwidth from a single DRAM is an advantage for networking, as it provides high bandwidth, but remains relatively cost-efficient because of low capacity. Consumer devices are much like networking cards, with the additional constraint that consumer devices are much more cost-sensitive. For instance, Sony’s PlayStation 2, which uses RDRAM, needs high bandwidth from a small memory capacity to reduce device count, mainboard area, controller pin-count (which reduces packaging costs and routing complexity), and thus total system cost.

High-end graphics cards, in contrast, require both speed and capacity. Modern graphics cards, such as ATI RADEON X800 XT and NVIDIA GeForce 6800 Ultra, carry 256MB onboard, the amount of memory that mainstream personal computers had in 2001-2002.

“Rambus has been in discussion with many different graphics processor manufacturers about XDR memory. There are not many choices for high-speed memory for GPU manufacturers, so it is natural that they would like to know about XDR and what it offers for their products,†the source noted.

XDR Compatible with DDR, DDR2, GDDR2, GDDR3

Even though Rambus’ memory devices have always been faster compared to available SDR or DDR solutions, customers in PC segment have been giving preference to cheaper mainstream memory products, especially keeping in mind that the difference in performance between RDRAM and competing solutions was not high. With graphics applications the things may change – in case visual processing units (VPUs) require extreme bandwidth, at some point XDR may become the only choice, especially if graphics processors’ developers may obtain compatibility between various memory types by using recently unveiled Rambus’ guidelines for memory controller.

Memory controller from Rambus provides support for mainstream DDR and DDR2 at speeds of up to 800MHz and graphics DDR, including GDDR1, GDDR2, and GDDR3 at up to 1600MHz data rates, Rambus DDR memory controller interface cells are full-featured drop-in physical layer (PHY) cells. The interface cells use proven technology that allow customers to improve time-to-market, minimize design risk and avoid potential re-spin costs. Rambus also offers system engineering services to further accelerate time-to-market, and ensure the interface operates at high frequency in the system environment. Rambus DDR interface solutions are ideal for a broad range of applications, from consumer multimedia and graphics systems to mainstream PCs and servers, the company said in late June, 2004.

Challenges on Rambus’ Path

Current GDDR3 that is used for high-end graphics cards nowadays may be clocked at up to 2.00GHz, providing bandwidth of 64GB/s for an application with 256-bit memory bus. The next incarnation of GDDR memory will be GDDR4 that is likely to be out next year and is expected to allow memory frequencies of up to 2.80GHz and effective bandwidth of about 89GB/s.

Today’s graphics processors require computing power more than memory bandwidth and developers of graphics chips will consider whether possibly more expensive XDR will deliver tangibly higher performance compared to GDDR4.

It is also not clear whether XDR supports 256-bit operation, but if it does, and with no substantially higher pricing, graphics processing units developers may bid on astonishing 102.4GB/s or higher bandwidth. In case it does not, 3.20GHz XDR devices are unlikely to become serious rivals for GDDR4 DRAMs able to work in 256-bit mode, a crucial feature for modern memory for graphics cards.

Officials for Rambus did not comment on the news-story.

Will this have some positive repercussions with regard to PS3?
Or since Sony only bought the ip and will fab it "themselves", it will have very little advantage for Sony?
 
This is interesting stuff, though not exactly console related. :)

It may or may not have some impact on what Sony's doing. Sony will stick to one density of ICs all through PS3's lifetime just like they do now with PS2, while PC graphics cards will increase in capacity over time. Thus it's not entirely certain Sony can benefit from increased supply and dropping prices from PC board sales should they NOT make their memory themselves.

I'm somewhat sceptical PC boards actually will use XDR, traditionally "nobody" ever used RDRAM on graphics cards, and I would think there is quite a lot of reluctance to change now. Memory controllers for standard DRAM-type memory is well advanced now and the same might not neccessarily be the same for XDR even if potential bandwidth is greater.
 
I would say no. I believe both ati and nvidia have had licenses for this tech for a few years . Same goes for amd. None of them have used it before
 
jvd said:
I would say no. I believe both ati and nvidia have had licenses for this tech for a few years . Same goes for amd. None of them have used it before

Right... but whose been the fab producing this 'tech' (XDR) for a few years? And, AFAIK, you're wrong concerning them holding liecenses for XDR.
 
Vince said:
jvd said:
I would say no. I believe both ati and nvidia have had licenses for this tech for a few years . Same goes for amd. None of them have used it before

Right... but whose been the fab producing this 'tech' (XDR) for a few years? And, AFAIK, you're wrong concerning them holding liecenses for XDR.

I was talking about rambus ram in general.

This isn't the only time that rambus has put out or announced fast ram. Remember the rambus for the p4 mobos was the fastest at the time. Amd could have gone that way but they went with ddr even though they had a lisence and still do to use rambus ram .
 
Jvd, you're right this has 'happened before', but you also have to realize that the parallel memory architecture used in today's RAM is running out of steam. The way GDDR3 connects to the memory controller isn't fundamentally different from the way the RAM in for example a C64 was accessed either. It still has separate row and column signals to address the memory, etc.

The serialized/packetized interface of XDR "currently" runs at 3.2GHz but scales to 6.4 in a FIRST step. GDDR3 runs out of steam much sooner than that.

It may well be we need something like XDR eventually to be able to continue scaling upwards in speed.
 
Can't really go by "Rambus in general." The tech changes, the price-to-performance changes, the quality changes... For ATi and nVidia, it would likely be "whatever makes the most sense financially" and on that point, we don't have anywhere near the figures they do to work things out. If it makes fiscal and competitive sense to pursue it, why wouldn't they? It's not like they'd have anywhere near the same concerns with it that regular consumers would buying RAM for their PCs.
 
Problem with Rambus Technology.

Latency is 50ns. Where as DDR Latency is about 2ns.

So... Unless XDR has solved this issue it may still not be a viable solution for Graphics systems. EXCEPT in cases where the GPU is using some large On Die Cache system.
 
I understand the diffrences between the techs.

I am saying this is not the first time rambus has had a fast ram tech and it had lost out once already. It may do so again .

That is why i don't see it happening.

I would think they would go to on die ram first if the gdr ram scales to slowly.
 
Hellbinder said:
Problem with Rambus Technology.

Latency is 50ns. Where as DDR Latency is about 2ns.

I think you might be going on partial truths here... you're comparing total latency for RDRAM versus one small part of the total latency equation for DDR.

http://www.aceshardware.com/Spades/read.php?article_id=5000186

There are some excellent articles over there on the subject.

One particular article (which I'm having a hard time finding) did some tests that showed RDRAM performing substancially better than DDR. I guess it turned out that it has a quicker turn around for reads followed by writes, where as DDR was much faster for sequential reads or writes. This could be the explanation for why graphics companies never used RDRAM in the first place.

Update: I found the article I was looking for... here's a link to the page talking about the turn around times...

http://www.aceshardware.com/read.jsp?id=50000346
 
Well, first of all, I think RAMBUS really should just die as a company due to their actions with SDRAM and JEDEC.

That said, I seem to remember that the major latency problem for RDRAM was due to the fact that only one chip on a RIMM is active at one time, for heat reasons. This is just an unfeasible scenario for graphics memory, as data is striped across all chips in such a way that they're really all active at the same time.

Secondly, we're really getting close to the physical limits for frequencies here. Before long, we'll need to move the memory chips closer to the die to keep frequencies increasing, meaning on-die or on-packaging memories. In the mid-long term, on-packaging/on-die memories may become much cheaper for low-cost designs, as the number of pins would be greatly reduced. In the long term, they may become faster as well, as this will alleviate problems the problems the physics brings with high frequencies (not to mention it may be cheaper to have wider busses....).
 
RDRAM doesn't have poor latency. When Intel was getting 250ns latency with their P4 chipsets Digital/Compaq/HP was getting 85-90ns from the same modules in their EV7 design (on die RDRAM controller). The reason Intel got poor latency was because they treated RDRAM as plain DRAM with a superfast interface.

Cheers
Gubbi
 
Gubbi said:
RDRAM doesn't have poor latency. When Intel was getting 250ns latency with their P4 chipsets Digital/Compaq/HP was getting 85-90ns from the same modules in their EV7 design (on die RDRAM controller). The reason Intel got poor latency was because they treated RDRAM as plain DRAM with a superfast interface.

Cheers
Gubbi

Even 85-90ns is VERY poor latency.
 
T2k said:
Gubbi said:
RDRAM doesn't have poor latency. When Intel was getting 250ns latency with their P4 chipsets Digital/Compaq/HP was getting 85-90ns from the same modules in their EV7 design (on die RDRAM controller). The reason Intel got poor latency was because they treated RDRAM as plain DRAM with a superfast interface.

Cheers
Gubbi

Even 85-90ns is VERY poor latency.

Very Poor is exagerating a bit don't you think? A P4 with DDR has a latency of about 100 ns depending on mem speed, cpu speed etc.

http://www.aceshardware.com/read.jsp?id=65000318
 
Moffell said:
Very Poor is exagerating a bit don't you think? A P4 with DDR has a latency of about 100 ns depending on mem speed, cpu speed etc.

http://www.aceshardware.com/read.jsp?id=65000318
P4 has an off-die memory controller (right?), and it gets around 70-90ns with PC3200 DDR. A64 has an on-die controller (like video cards, right?), and it hits 40-50ns with PC3200 DDR. I'm thinking the A64's much lower latencies factors into its current (total) dominance in game benchmarks. The question is, will an on-die controller still yield much greater latency on XDR than GDDR3? And, if so, will that affect scores enough to make XDR a bad choice until it can compensate with overwhelming bandwidth?

(I haven't the first clue about CPU and gaming performance, though, so it's quite possible--likely, in fact--that latency is a minor factor.)
 
Pete said:
Moffell said:
Very Poor is exagerating a bit don't you think? A P4 with DDR has a latency of about 100 ns depending on mem speed, cpu speed etc.

http://www.aceshardware.com/read.jsp?id=65000318
P4 has an off-die memory controller (right?), and it gets around 70-90ns with PC3200 DDR.

And you'll note the 256 byte stride comment. What this shows is the latency the processor sees when prefetching is active. Random access latency is still around 200 ns. Athlon XP's are aroung the 240-280 ns mark

Cheers
Gubbi
 
Ah. Are there benchmarks online that show random access? I just checked TR's 3D graphs.
 
T2k said:
Even 85-90ns is VERY poor latency.

You are joking right? EV7 does ~85nS with untweaked 800 Mhz DRDRAM. Using the timings available with current DRDRAMs the latency would be ~75nS. Which is VERY VERY good.

Aaron Spink
speaking for myself inc.
 
Back
Top