Tony and Thread - Byte.com column on RAMBUS and PC memory
byte.com
====================================================================
No Summer Dog Days For Chips This Year
Direct Rambus DRAM May Be Summer Blockbuster.
By Mark Hachman
June 21st, 1999
In This Article
• No Summer Dog Days For
Chips This Year
• I Feel The Need: The
Need For Speed!
• Waterworld Wasn't Cheap,
You Know
• Is El Camino Real?
• The Balcony Isn't Closed
Industry wonks used to refer to the summer months as the "summer slowdown," when all the marketing types jetted off to exotic locales and let journalists camp out for the latest flicks.
Not this year.
During June, July, and August, chip vendors will be working frantically to prepare the industry for its own blockbuster: Direct Rambus DRAM. Like any good feature film, the launch of Direct RDRAM has been seeded with months of hype, much of it created within the industry. But the question for many will be whether the initiative will prove to be a success of Godzilla-like proportions, or merely another phantom menace.
Trying to compare Direct Rambus DRAM and its nominal competitor, PC133 synchronous DRAM (SDRAM), is fraught with perils akin to the opening scenes of Saving Private Ryan, a nightmare of deafening public-relations outcries and explosive technical discussions layered over the cruel, uncaring economics of chip manufacturing.
At Last We Shall Reveal Ourselves To The Jedi
Main memory within a PC has evolved over time, through different evolutionary technical specifications: from fast-page-mode, through extended-data-out (EDO) and finally to synchronous DRAM, or SDRAM. Each flavor has had its own technical selling points, but in general each has transferred data faster or with shorter delays than the prior generation.
Along the way, a host of alternative-memory types has been proposed, such as Enhanced SDRAM, Direct Rambus, and Virtual Channel Memory, to name but a few. Direct Rambus is the third generation of Rambus parts; prior Rambus generations are used within the Nintendo64 game console. But each of these memories has claimed to be revolutionary, not just evolutionary. For Rambus, that's the promise and the problem, all in one sentence. A revolution can produce a dramatic change, but requires both material and manpower together with a political, social, and economic commitment. Shifting the PC to Direct Rambus DRAM requires the support of DRAM, chip set, OEM, and test-equipment manufacturers, each from a technical, economic, and, yes, even a political perspective.
Unfortunately, the commitment to Rambus seems more like the attitude toward Vietnam than to World War II, no disrespect intended. And the debate over manufacturing and selling Direct Rambus parts has provoked fusillades of press releases, clouds of FUD [Fear, Uncertainty, and Doubt], and even a lawsuit, later retracted.
On the surface, the issue seems fairly clear-cut. Direct Rambus DRAM is designed to transfer data at speeds of 1.6 gigabytes per second at "800 MHz," far faster than the 800 megabytes of data that PC100 SDRAM moves during the same period. PC133 increases the data bandwidth to 1.06 gigabytes per second.
Some analysts have also pointed out that Direct Rambus intrinsically generates more latencies, or delays in accessing data; others have said this could be negated through fixes like optimized software code. Of course, it isn't really that simple. Direct Rambus DRAM doesn't really run at 800-MHz; it transfers data 800 million times a second, twice per clock cycle. The chip really runs at half the rated speed: 400 MHz for an "800-MHz" Direct RDRAM. There also isn't one speed of Direct RDRAM, but at least three: "600-," "700-," and "800-MHz" versions.
Intel's taken the additional step of requiring suppliers to call a "800-MHz" Direct RDRAM a "PC800 RDRAM chip", to provide "consistency with the current "PC100 SDRAM' naming," Intel's confidential documentation says.
(I'm content to let the market decide the Direct Rambus/PC133 issue, and in the interim both sides have taken some cheap shots at the other. But to me, labeling the distinctly different Rambus architecture with a SDRAM label is, in effect, taking an orange and painting it apple red. It also seems consistent with Rambus' unspoken policy not to let its licensees speak openly about its technology's disadvantages -- a practice Rambus admittedly seems to have relaxed.)
One DRAM vendor summed the debate up quite nicely. "Rambus is a nice, fast, sexy device," Farhad Tabrizi, director of strategic marketing for memory products with Hyundai Electronics America's semiconductor division, told Electronic Buyers' News. "If you want performance, you have to pay for it. That's the bottom line."
And for a BYTE reader, that's probably the bottom line, too. With a larger die size, Direct Rambus DRAM fundamentally will cost more to produce on manufacturing processes with equivalent line-widths, and DRAM vendors have reported all sorts of difficulties manufacturing the parts. They have also complained about the dearth of Rambus chip and module testers, which adds to the processing cost.
According to Tabrizi, a 128-megabyte Direct RDRAM memory subsystem is estimated to cost between $200 and $250 this year and could drop to less than $150 by the end of 2000.
Using Pricewatch.com, a website for purchasing hardware components online, one can buy 128 MBs of PC100 SDRAM today for under $80. Tacking on an estimated 33 percent premium to that price for PC133 speed -- an arbitrary choice, and a figure that isn't necessarily justified given the precipitous price drops even in the latest memory chips -- generates an estimated price of $106 for 128 MBs of PC133 SDRAM.
Given those admittedly very rough estimates, PC800 Direct Rambus seems to outperform PC133 SDRAM on paper by about 51 percent, but could be priced 89 percent higher at the outset. If it's true that PC600 Direct RDRAM transfers data at 1.2 gigabytes per second, then it would outperform PC133 SDRAM by only 13 percent, but be priced 25 percent higher. Note that Tabrizi's example refers to the price of Direct Rambus DRAM later this year, while PC133 SDRAM is available for purchase now. (At some point, I'm sure you'll be able to purchase Direct RDRAM online, as well.)
For a BYTE reader, the debate essentially stops there. Unless Direct RDRAM is proven to be a money-losing proposition for DRAM vendors -- a possibility, but one that Intel has helped alleviate by making substantial investments into DRAM manufacturers -- Direct Rambus will come to market, most likely in September. If supply is tight, the chips will be restricted to the most expensive machines, and even possibly to a certain number of OEMs. Because PC600 Direct RDRAM will be the easiest to manufacture, it's probable that Intel will attempt to use them to seed the PC market and create both demand and an upgrade path to the higher PC700 and PC800 speeds.
The issue of the delay in the Intel 820 or "Camino" chip set is significant, but more so for OEMs manufacturing PCs based upon the new memory type than end users.
Camino, the first chip set to include the Direct Rambus interface, is scheduled for a revised release date in September. OEMs could thus be hard-pressed to ship Direct Rambus-enabled PCs by the holidays, possibly affecting PC sales. Historically, consumers have waited to purchase PCs if they knew something really good was not far off. Incidentally, if anyone can tell me the combined effect of the Camino delay and Y2K PC purchasing patterns will have on the economy, please do.
It's also possible that Intel will finally just pack it in, give it up, and use PC133, a stance the company has publicly, flagrantly avoided, but privately admitted, customers say. Intel CEO Craig Barrett has already publicly acknowledged the presence of a DRAM contingency plan. It's an issue of marketing, with no real significance to the end user.
Though it's true that Intel has publicly backed Rambus to the hilt, it's not alone. Some DRAM vendors say they hope that Direct Rambus DRAM will let them sell high-speed chips with high-powered margins.
It's also very interesting to note that two of the chip sets prominent manufacturer Acer Laboratories has on its road map "support Direct Rambus DRAM," its customers say. Both the high-end Aladdin K7 II and III will support AMD's forthcoming K7 microprocessor.
In the hierarchy one customer presented, the Aladdin K7 I will ship in the third quarter, with 100-MHz frontside bus support for PC133 SDRAM. By the third quarter of 2000, the customer said, the Aladdin K7 III would replace, not complement, the K7 I with a 133-MHz frontside bus and Direct RDRAM support. Acer Labs' Aladdin Pro III for the Pentium II and III reportedly includes PC133 support, but the company is still awaiting a P6 bus license from Intel.
According to Dirk Meyer, vice-president of engineering within AMD's Computational Products Group, AMD will not initially support Rambus in its supporting K7 chip sets, but plans to later.
Meanwhile, information surrounding chip set maker Via Technologies' K7 plans is not known, but the Apollo Pro II, scheduled after the first year, includes support for a "new DRAM technology" which might be Direct RDRAM.
Will Via support PC133 SDRAM? It would like to, but as Forbes first pointed out, Intel sent a letter to its customers advising them that its PC133 SDRAM chip set, the Apollo Pro133, isn't covered under Intel's patent license. That followed a patent-infringement lawsuit against Via that Intel said it filed in error. Via executives were more suspicious. Since Intel won't let Via ship the Pro133 for now, Via's PC133 plans have been put on hold.
Well, then: thumbs up or thumbs down on Direct Rambus? I don't know. With DRAM prices continually falling, it seems silly to spend the extra money on memory, even higher-performance memory.
On the other hand, DRAM vendors, analysts, and I can merely speculate what Direct RDRAM will cost, and what its performance benefits will be. Testing of the first round of Camino, Direct RDRAM-based PCs will be very telling.
What I suspect will happen is this: Direct Rambus DRAM will penetrate the market in much the same manner as Intel's own microprocessors have done, firmly controlling high-end PCs, but losing a heated dog fight in PCs selling for less than $1,200 as long as PC100 or PC133 SDRAM is manufactured. While AMD's success as Intel's chief microprocessor rival is tied to its single fab, the swarm of DRAM vendors can quickly shift to high-volume PC133 production if the market demands it.
So sit back, relax, and wait for Direct Rambus. Grab some popcorn. Save the Rambus Inc. Web page as a screensaver, if you'd like. Just remember that a movie can be quite different than the trailer.
======================================================================
PB |
