Rambus takes memory bandwidth lead

MEMORY SPEED was the hot topic until mid year, with a flurry of advances in the past couple of years, going from SDRAM to DDR, and DDR itself stepping up from 200 to 400 in a quick succession of steps. When DDR advances petered out, both Intel and now AMD went to dual channels, and are looking forward to DDR2 as the next saviour. Therein lies the problem, DDR2 is still a few months away.

Dual channel DDR-400 has been with us for most of this year, and by the time it is supplanted, it will have been about a year with no memory advances. Gasp, what will people like me do without minor speed increases to crow about? Well, if you read the INQ regularly, you will know we whine about it.

Sadly, the engineers at SiS and Rambus are doing something about the problem, and we won't get our chance to whine. SiS has just introduced the highest memory bandwidth chipset out there, and we are not talking baby steps either. Since the technologies are different in many ways I won't compare them using MHz, but bandwidth.

DDR200 is also known as PC1600, offers 1.6GBps of bandwidth, 200 MHz * 8 bytes wide = 1600 MBps. As DDR moved up to DDR400 in 66MHz steps, bandwidth went up to 3.2GBps, or 6.4GBps with 2 channels. By the time DDR2-533 hits next year, the bar will be raised to a mere 4.3GBps, or 8.6GBPs in dual channel form.

If that sounds like a lot, the new SiS R659 coupled with 4 channel 1200MHz RDRAM will have an aggregate bandwidth of 9.6GBps. We hope to get a review sample before the end of the year, with general availability shortly after that. With any luck, before Christmas, Rambus will have raised the bandwidth bar higher than DDR2 is set to achieve in all of 2004.

There are three questions that need to be answered, latency, price and availability. Price and availability go hand in hand, and are kind of a chicken and egg problem. It will surprise no one that since Intel unceremoniously dumped Rambus at the altar, the memory modules are getting harder and harder to find, and more expensive to boot. Meanwhile, they have been getting steadily faster, 1066 RDRAM is no trick anymore, but a quick check of the local shops shows only a handful of sticks, dwarfed by masses of DDR modules. This is not a good sign for availability of DRDRAM 1200, but since Samsung and Elpida are making the chips, there should be no major shortages.

With luck, the boards, including a rather slick looking Asus P4S13G with it's 2 right angled RIMM slots should be here around Christmas, and the memory to plug into it should be out at the same time. No word yet on price, but I would expect at least a moderate premium over DDR. If the numbers shown to us prove accurate, a 10-15% increase over the Intel i875, it may well be worth a few more dollars.

The more important thing is that the slides shown to the Inq have DRD/1600 all over them, but no work whether the chipset supports it or not. It makes you wonder how much memory bandwidth the vaunted P4 can handle before it starts to gag, and how much it would cost to hear that noise. We should know in just over a month.

The more pertinent question is latency, long considered the achilles heel of RDRAM. Some things make RDRAM sing, others make it fall down with a loud squishy thud. Those squishing sounds are usually accompanied by latency sensitive apps. The first good thing is that with increased clock speeds comes decreased latency, in absolute time, if not in clock cycles. If nothing else, this will bring a large latency drop over the more common DRD/800.

The other technologies that should help are prefetch and on chip caches. The memory controller in the R659 does both, and if it works as promised, it should lead to much lower latency in most situations than older DRDRAM implementations. The numbers we have seen show that the R659 has latency times no worse than DDR400, but can be several times better under ideal circumstances.

Overall, in a month or so, when the numbers trickle out, we will have definitive answers to all these questions. Right now, things look good for the SiS chipset, but as usual, introductory presentations tend to color things in a good light. Rambus did for the longest time have the fastest solution available for the P4, but then politics and lawyers got into the game and ruined it for all of us geeks. With the lawyering subsiding, and politics out of the way, I wonder if this could signal a Rambus resurgence. Stay tuned, it will be interesting at the very least. µ