Commentary on Rambus and RDRAM
Oct 02 '00
First, a little history. Obviously there was a need to develop a new RAM standard. SDRAM has been the predominant memory technology for quite some time. It has definitely served its purpose well, as it's in the majority of PC's out there in the world today. While SDRAM was a viable technology for a long time, it's my feeling that it may be holding back the industry now more than ever, and consumers' attitude towards competing memory technology may be hurting the industry.
You see, people seem to romanticize with memory. This has been a problem for years now, dating back to my college days 10 years ago when 2MB of RAM would cost you well over $1000. Back then, everyone ran DOS, very few people ran Windows, and even fewer ran Windows 386. 386 PC's cost somewhere in the neighborhood of $5000 and came with memory barely adequate to run stripped-down versions of "Windows Runtime", packaged with many GUI applications. In this time, it was established that the most important component in your system was memory, and having as much memory as possible was the only sure-fire way to get adequate performance out of your PC. 10 years ago, the industry was converging on several new technologies. Windows was maturing, Microsoft and IBM were working on OS/2, the proposed "future" at the time. All these new technologies were going to offer new functionality which in turn would require more memory. Of course the memory industry knew this, and thus, the price remained high.
Memory manufacturers justified memory costs in many ways over the years. They knew people would want more memory for these new converging technologies, so they inflated prices in any way they could. How many times have YOU heard the rumor of some random chip manufacturing plant in a random asian country burning down, causing the huge surge in prices? Over the course of 10 years, I have heard no fewer than 3 rumors of memory plants burning down, sometimes causing near-200% increases in memory prices. All these rumors have been false, to the best of my knowledge.
While several more factors aside from RAM play a role in performance today, this still remains true. More RAM usually means better performance, despite the platform. Whether you were populating boards with chips, snapping in SIMM or SIPP modules, or popping in SDRAM modules, you were almost guaranteed a boost in overall performance by adding more memory if your system was bottlenecked there.
Fast-forwarding to more recent times, Intel remains the predominant player in the PC market with 80% of the market cornered. PC's are now everyday tools, rather than an expensive luxury. We're living in times where processor speed now doubles once every 18 months (Moore's law). The internet has spawned entire communities, linking people with technology in ways no one predicted (well, except for maybe a few science fiction writers). Memory technology is now approaching a time where it's going to have a hard time keeping up with faster processors.
Intel saw this coming.
Intel felt that SDRAM had far too many limitations to be considered as a viable long-term technology. There's still quite a bit of uncertainty in certain circles as to what the future holds for memory standards, but a little intellectual property company called Rambus is attempting to develop standards that satisfy future memory throughput needs.
Now, Rambus seems to be a company everyone in the PC enthusiast community loves to hate. Why is this? I can see several reasons. Some true, some blatantly false, some unconfirmed. I will try to separate fact from fiction as best I can.
One of the primary complaints (if not THE primary complaint), is the price of RDRAM. Yes, for the AFTERMARKET buyer, RDRAM is expensive. For the corporate volume buyer, no. Early figures claim that the increase in cost for an entire RDRAM-enabled machine was as little as $100. When RDRAM was first released several months ago for the PC platform, you could expect to pay $1200 for a 128MB PC800 ECC RDRAM RIMM module. Why? Limited initial production made RDRAM availability VERY limited, and much RDRAM as possible was allocated for large OEM's such as Dell and Compaq. I'll sum it up this way: for every RDRAM RIMM they sold to an aftermarket buyer, that was one less system Dell and Compaq could build and sell to an end-user, with which to make a decent profit.
Remember, they're all in this for the money.
A lot of people have talked of the poor yields memory manufacturers are experiencing with RDRAM. Fact is, with any new technology, refinements in the manufacturing process have to be made. Yields for RDRAM have improved dramatically over the course of the past few months. This has both improved availability and driven the cost down. What cost $1200 a few months ago costs roughly $250 now. ONE FIFTH the price it was. If this trend continues (of course, it will), RDRAM will be affordable to the end-user sector in a matter of months. As the refinement process continues, prices will drop as yields improve.
A lot of people claim DDR SDRAM is a much better solution than Rambus RDRAM. Maybe as a short term solution, but there are several key facts killing DDR SDRAM as a viable technology. This has been the predominant reason there is "uncertainty" in the memory standards game. The first problem with DDR SDRAM is its tardiness in the market. While VIA has been attempting to fan the fire for DDR SDRAM, there still isn't a production DDR SDRAM solution for PC's on the market. Meanwhile, Intel is already shipping 820 and 840-based solutions to OEMs, and has been for some time. In fact, I have this nagging feeling that they'll be releasing the Pentium 4 before any manufacturer releases a DDR-SDRAM solution. This gives Intel a jump on any competition that could come from VIA or AMD. While enthusiasts love to hug the underdog, business IT departments don't. Regardless of what enthusiasts think, it's business that drives the market, not necessarily the opinions of Doctors-turned-hardware-gurus who make their livings by posting sensationalistic journalism on their webpages.
Another problem with DDR SDRAM is its own future. DDR SDRAM is fast, without a doubt. But it has its limitations. DDR SDRAM is basically double-pumped SDRAM. The increase in transfer rate is basically accomplished in the same way RDRAM's transfer rate is accomplished at one level. Sure, you can make it as fast or faster than the current crop of RDRAM RIMMs available. But then what? What's its future? If DDR SDRAM is already a bottleneck for video applications (as seen by the newest Nvidia Geforce 2 GTS card), how will it keep up when processors reach 1.5ghz late this year? If you were Intel, would YOU base your future on a memory technology that is both currently unavailable and has an uncertain performance potential in the future?
Rambus is already developing an RDRAM standard that accomplishes 17GB/sec throughput out of a single Rambus memory controller. Using dual Rambus channels, that number is effectively doubled. I'd say that's a significant improvement over the 1.6GB/sec available now (3.2GB a sec if you're using dual Rambus channels a-la the 840 chipset platform). Intel had no choice but to bet its future on Rambus technology. They saw that these kinds of transfer rates would become a reality, and that they would satisfy their processor speed needs. Basically, there was really no other choice for Intel.
AMD is now finding themselves in the same boat. They have big plans for future processor families, and they too are going to find themselves under the gun. Will they go with DDR SDRAM? Maybe. They certainly did some homework on the 760 chipset they're developing, but will DDR SDRAM see the light of day? All Rambus REALLY has to do is claim that Micron can't release it because it's pending litigation thanks to the new flurry of lawsuits. That could push DDR SDRAM's release back even further.
If you read the "writing on the wall", however, you may find that their choice might be Rambus. What makes me say this? Simple. Go to careers.yahoo.com and do a search in California for AMD job listings. Limit your search to the keyword "Rambus". That's right. AMD wants Rambus-savvy engineers to work on SOMETHING. What they want them to work on is anyone's guess. It's my GUESS (read again: GUESS) that they want to INVESTIGATE the long-term potential for including Rambus technology in their product line.
The Tomshardware.com fiasco
So how about those Benchmarks? I'm sure you've read tomshardware.com's jaded benchmarks showing that Rambus-based systems underperform when compared to BX and VIA counterparts.
Let's dissect that for a moment.
First of all, they tested the VIA platform, the 820 platform, the 840 platform, and the 440BX platform. All platforms were run at their native speeds with the exception of the BX platform, which was running at an overclocked frontside bus speed of 133mhz (the BX was NOT designed to do this, but hey, tomshardware always glamorized overclocking as being the sure-fire way to make your system perform at its peak, a sentiment I wholly disagree with coming from the business world). They make note of how "Latent" RDRAM is, yet never really show how this is Rambus' achilles heel. I don't think anyone has really proved just how much of an issue this is. I think people just "assume" the problem is there.
They claim that RDRAM RIMM modules get "very, very hot very, very quickly" (can *anyone* confirm this? The RIMM modules in my machines at work and at home barely get warm).
In their testing, the following benchmarks are apparent to me:
The overclocked BX board is fastest, though running WAY out of spec.
The i840 takes a close 2nd place with its dual Rambus channels
The VIA133 and Rambus systems basically alternate between 3rd and 4th place
Dead last, the i820 with a memory translator hub is the slowest solution and a big mistake (the one "fact" I agree with).
So what is Tomshardware proving exactly? That an older overclocked setup is the way to go? That VIA's PC133 is the way to go? How is an overclocked BX system beating a non-overclocked Rambus system "embarrassing for Intel"? Just how many people do you expect will be running a BX at 133mhz on the front-side bus? How stable could it possibly be? Even if their 815 is faster than the 820, how are you going to use PC133 SDRAM on a motherboard that sports a 400mhz FSB (such as the Pentium 4's)?
Tomshardware, how about answering these questions:
Why didn't you include non-overclocked speeds when testing the 440BX chipset to compare how these chipsets operate in-spec?
If that's "not your style", why not include benchmarks of overclocked RDRAM and VIA-based system tests? How do they scale in comparison?
My thoughts on the differing memory technologies
In conclusion, I'd like to offer the following OPINIONS.
1. Rambus offers potential that no memory technology currently offers. Granted, the 820 chipsets of today may not fully utilize its potential (most first-cut technologies don't). In time, Intel (and possibly AMD) will develop solutions that more effectively use what Rambus-based technology has to offer.
2. SDRAM is a dead-end. Intel knew it. Sure, you can double-pump it (DDR SDRAM). You may even find out a way to quadruple-pump it (I read something somewhere about some random company developing this, I don't recall who). But before too long, you're going to hit a dead-end with it. It's too late to market. IBM may pledge support for it, but if you ask me, they're just covering their asses. Fact is, IBM already markets and sells RDRAM-based systems. Go figure.
3. Rambus' intellectual property may keep other companies from developing future memory technologies to compete with it. Rambus holds more than 100 patents in their Intellectual property portfolio. Truth be told, Rambus doesn't manufacture anything. They develop specifications. They are a pure intellectual property company. They collect 1-2% royalties on whatever memory technology that utilizes any of Rambus' patents. All they do is sit back and collect money. Great business model, huh? You patent-lawsuit haters might cringe at Rambus for the hitachi lawsuit (which they settled out of court, and now pay licensing fees to Rambus), but the bottom line is Rambus doesn't care if you manufacture a competing memory technology as long as you keep your agreements with Rambus. Hitachi made a mistake in my opinion. They used a former relationship with Rambus to develop a competing technology and landed themselves in court. Micron and Infineon are bellyaching over the royalty now too, but in the same time, large semiconductor companies (like NEC) are signing on. Take a look at Micron's market cap since they announced the lawsuit. My prediction here is that they'll settle out of court after being pressured by their shareholders to do so. This lawsuit adds uncertainty to micron's bottom line, and shareholders don't like that.
4. Memory manufacturers will continue to improve yields, and as the popularity of the 820 increases and the willamette (Pentium 4) gets close to shipping. RDRAM will be become available and affordable to the aftermarket. It's already dropped to 1/5th of what it was at launch. At launch, a 128MB RDRAM module cost $1200. You can now get that same amount for $250 from pricewatch. There is absolutely no reason to believe this pricing trend won't continue. My estimate: By the time DDR SDRAM makes it out the door (assuming it's not prevented by pending litigation), pricing will be comparable between the two.
At this point, all roads lead to Rambus, whether you like it or not.
 |
|
|
|
|
|
|
Epinions.com ID:
kweckstrom
|
 - Top 500 |
|
Member: Karl Weckstrom
Location: Little Ferry, NJ
Reviews written: 100
Trusted by: 99 members
About Me:
Voted "Most Likely to be Photographed as a Bigfoot Sighting" by his senior class.
|
|
|
© 1999-2009 Shopping.com, Inc.