Dividing (and Conquering?) demand -- It's not crystal clear which of the new DRAM technologies will dominate.
John Day October 11, 1999,
The team looks great on paper, but paper isn't where fortunes are won or lost. Intel Corp. and Rambus Inc. promise that in the next few years, Rambus will be the mainstream memory technology, having unseated the now-reigning PC100 and its follow-on parts, PC133 and DDR SDRAM.
Others aren't so sure.
What makes Rambus so formidable is its system-level architecture and the fact that it has Intel's commitment. But delays in the launch of Intel's Rambus enabling 820 chipset, Camino, will seriously affect the future of Direct Rambus DRAMs.
It's uncertain, however, what bearing this will have on competing technologies. DDR does not have comparable memory controllers. In fact, Intel has only recently announced that it will build a new memory controller chipset to support PC133.
That decision was based on requests from customers and on Intel's perception that memory-chip makers can yield PC133 devices from PC100 die, according to a spokesman for Intel, Santa Clara, Calif.
Via Technologies Inc. and others will help flesh out a high-performance, low-cost PC133 platform, analysts said.
Taking a leap
Chipset issues notwithstanding, there's general agreement in the industry that the Direct RDRAM's time has yet to come.
"Today's applications are designed for today's technology, and they work fine," admitted Jeff Mitchell, business development manager at Rambus, Mountain View, Calif. "Rambus won't make a lot of difference. It will be like putting a supercharged race car on a freeway. It can only go as fast as the freeway and the other vehicles on the freeway will allow."
Rambus will enter the memory market at the high end and work its way down, Mitchell said. "The real benefit of Rambus is that it allows headroom for future applications and technology and eliminates the memory system as a potential bottleneck," he said.
Direct Rambus DRAM represents a major transition in memory technology, the Intel spokesman said. "As happened during the transition from fast-page mode to EDO, and from EDO to SDRAM, there will be a period of a couple of years during which the previous technology exists with the new one."
Rambus is expected to capture more than half the DRAM market by 2002 or 2003, and DDR will account for most of the balance, according to Steve Cullen, an analyst at In-Stat Group, Scottsdale, Ariz.
By 2002, Rambus will control about a third of the memory market, predicts Avo Kanadjian, senior vice president of memory marketing at Samsung Semiconductor Inc., San Jose. He sees the other two-thirds split between DDR and PC100/PC133 SDRAM.
Serving different segments
"It's a three-horse race: PC133, Rambus, and DDR," said David Bondurant, vice president of marketing at Enhanced Memory Systems Inc., Colorado Springs, Colo. "They will serve different segments because no one type of memory can serve all applications." (See related story on page E3.)
"It looks now like DDR and Rambus will coexist," added Will Mulhern, product marketing manager at NEC Electronics Inc., Santa Clara. "SDRAM will be important in 2000, significant in 2001, and still there in 2002."
Mulhern estimates that as much as 70% of the market will stick with SDRAM next year, with that share declining to 30% to 40% in 2001. The PC100 will continue as the mainstream memory device next year because of its value at the low end of the PC market.
There is likely to be substantial overlap between generations and technologies.
Although some predict Rambus will eventually dominate, it won't replace SDRAM entirely.
"SDRAM will work just fine in many applications, especially in very large systems with wide data buses," In-Stat's Cullen said.
PC100 SDRAM chips operate at 100 MHz on a 64-bit (8-byte-wide) data bus to achieve throughput of 800 Mbits/s. PC133 SDRAMs operate at 133 MHz for throughput in excess of 1 Gbit/s. DDR chips, at double the data rate, offer throughput in the 2-Gbit/s neighborhood.
Rambus devices feed data at 800 Mbits/s, but are expected to drive two or more channels for throughput of 1.6 Gbits/s and beyond.
"It's easier to add Rambus channels than it is to widen a DDR bus," said Kevin Kilbuck, memory applications engineering and technical marketing manager at Toshiba America Electronic Components Inc., Irvine, Calif.
Advanced Memory International Inc. (AMI2), the consortium formed to promote DDR technology, anticipates that DDR2 with 4-Gbit/s throughput will be in production prior to 2002. But at the moment, the capacity for performance improvement appears to be greater with Rambus than with DDR.
Tough comparisons
"In the past, memory technology has been very compartmentalized," Rambus' Mitchell said. "DRAM companies did a good job of making DRAM chips, but it's been up to the systems makers to put all the pieces together.
"Rambus took a different perspective in looking at the whole system. We're going to see processors approaching 1 GHz, and memory systems have to be able to run faster to keep pace, but that's not going to happen simply by making memory components run faster. That's the easy part," Mitchell said.
Device makers counter that, until now, the best way to increase PC performance was to add memory. "You could get a huge increase in system performance by doubling the amount of system memory," said Jeff Mailloux, director of DRAM marketing at Micron Technology Inc., Boise, Idaho. "By comparison, doubling the speed of the processor didn't mean as much."
Rambus has an advantage in performance, but a disadvantage in cost, noted Bob Fusco, marketing manager for the DRAM division of Hitachi Semiconductor (America) Inc., San Jose. The price premium for Rambus could be as much as three or four times that for SDRAM, and the performance improvement will be slight by comparison, he said.
"The whole notion of Rambus is based on bandwidth being the most important determinant, but latency-the time required to get the first bit of data-is equally important," Fusco said.
Megabytes matter more than megahertz. "Running back and forth to the hard drive to get files drags system performance way down," Fusco said. "It's the amount of memory, then latency, then bandwidth that determines system performance."
Mitchell counters that the latency of Rambus devices is the same as that of the PC100 because both devices use the same memory core. The memory core dominates latency; the interface only makes a small difference, he maintains.
In the mainstream
At the low end, the priority has always been cost.
"And by working to drive costs down to the lowest possible level, we were responding to our customers," Micron's Mailloux said.
That kind of cost reduction makes the DRAM market work, said analyst Bert McComas at InQuest Market Research, Gilbert, Ariz. DRAM is the poster child for economy of scale, he said, adding that the cost per bit of memory must continue to go down.
Devices are either ramping up to or down from the mainstream level; there is no other alternative, he said.
"PC100 is currently the mainstream, and PC133 is obviously what's going to ramp up next," McComas continued.
But PC133 users are hard to find because Intel hasn't yet launched a 133-MHz frontside bus, McComas said.
"The PC133 performs magnificently and is a generous upgrade," he said. "It's adequate for a year or 18 months. It's the same amount of upgrade that the PC100 was over PC66, or the PC66 was over EDO. It's completely in line with the his-tory of the universe."
Rambus and DDR are both ahead of their time, according to McComas. For now, they make sense only in specialized applications in which price premiums are palatable. One or the other will become mainstream when the market is ready to take advantage of its strengths, "but the probability of DDR going mainstream is a heck of a lot higher than it is for Rambus," McComas said.
Which way to go?
Availability is perhaps as critical as cost.
"There are certain DRAMs we have today that customers would like to continue using. But if the mainstream is moving, those customers have to be concerned," said Cecil Conkle, assistant vice president of DRAM marketing at Mitsubishi Electronics America Inc., Sunnyvale, Calif. "Intel has such a powerful and dominant position in CPUs and chipsets that they can have key input and may even dictate the way technology is going to go. And Intel has selected Rambus."
Achieving the high speed promised for Rambus with acceptable yields requires 0.18-micron or smaller line geometries, and the DRAM industry isn't ready for that, Conkle said. Rambus has "a considerable die-size and cost boundary," although that obstacle becomes less severe at higher densities.
The lack of high-speed test equipment is another problem device makers must overcome if they are to participate in the Rambus market. And participate they must.
"Memory-device makers have to offer a wide selection, or OEM customers won't see a firm as a strategic supplier; but it takes a lot of investment to have all three available," said Jan du Preez, vice president of memory products at Infineon Technologies Corp., Cupertino, Calif. "Our preference would be to focus on a specific niche, but we're supplying PC133 now, DDR is on our roadmap, and we will have Rambus available.
"OEMs are struggling to predict their supply and demand models. If we start focusing on one application, or on making a specific product, we'd have a lot of capacity available," he said. "We have to have a wide spectrum of products or we won't be on our customers' radar screens. We have no choice."
-John Day is a freelance writer based in Stone Mountain, Ga.
---
Comparing The Chips
Speed
- PC100 SDRAM chips operate at 100 MHz on a 64-bit (8-byte-wide) data bus to achieve throughput of 800 Mbits/s
- PC133 SDRAMs operate at 133 MHz for throughput in excess of 1 Gbit/s
- DDR chips offer throughput in the 2-Gbit/s neighborhood
- Rambus devices feed data at 800 Mbits/s, but are expected to drive two or more channels for throughput of 1.6 Gbits/s and beyond
Latency
- The latency of Rambus devices is the same as that of the PC100 because both use the same memory core
Price, applications
- The premium for Rambus could be as much as three or four times that of SDRAM
- Rambus and DDR make sense only in specialized applications in which users are willing to pay higher prices
- PC100 is currently the mainstream chip, and PC133 is ramping up
techweb.com |
