Will New Architectures Provide A Path To Recovery? -- The trail is still tortuous for memory-module suppliers, which are grappling with new design and test challenges.
John H. Mayer
Although hope is building that DRAM ASPs may have finally bottomed out after a couple of extremely tough years, memory-module suppliers are hardly out of the woods.
The coming transition to new high-performance DRAM technologies promises to winnow out weak competitors and consolidate the market, according to analysts.
"It's going to take a considerable up-front investment and extensive engineering expertise to survive," said Sherry Garber, an analyst at Semico Research Corp., Phoenix. "Module suppliers can't just be assembly people anymore; they have to understand what they're assembling and where it's going."
Fueling concern is the coming migration from SDRAM to Direct Rambus and double-data-rate (DDR) devices. Supporting data rates up to 10 times faster than current solutions, these new technologies will open up design issues that module manufacturers have never dealt with before.
"You have to spend extra money on your machines, your vision systems, and the actual printing of the paste on the printed-circuit board, not to mention deal with the PCB issues," said Adrian Proctor, general manager of U.S. operations at memory-module supplier Dane-Elec Corp., Irvine, Calif. "Practically every DRAM manufacturer is going to have a slightly different pin layout, so goodness knows, it will be chaos for a while."
Adjusting to a new infrastructure
The Rambus architecture's memory-bus speed of 800 MHz creates new design requirements, such as the use of spread-spectrum clocking to deal with EMI, as well as the introduction of new types of connectors, said Boba Popovicv, director of engineering for Irvine-based module supplier Camintonn Corp.
"For Rambus, we're adding a packet transition, a voltage transition, and a density transition in a wide variety of module form factors, as well as changing the DRAM interface, so it's becoming quite complex to cover the entire matrix and supply those to customers," said Bill Johnston, vice president of strategic marketing in the memory group at Smart Modular Technologies Inc., Fremont, Calif.
Memory-module suppliers give Rambus Inc. and Intel Corp. much of the credit for quickly bringing into place the infrastructure for the new technology. To ease the industry through the transition and accelerate time to market, Mountain View, Calif.-based Rambus in September announced a component-validation program similar to what Intel used with PC-100 SDRAM components. The program establishes uniform verification procedures for Direct Rambus components, including Rambus-in-line memory modules (RIMMs), and promotes compatibility and interoperability across various suppliers' products.
Initially, Rambus will direct the verification process while enabling independent test houses to support the procedures. Eventually, suppliers will be able to "self-validate" their components.
Meanwhile, suppliers have brought to market not only Rambus chips, but all the ancillary components. AMP Inc., Berg Electronics, and Molex Inc. have announced RIMM sockets, while Cypress Semiconductor Corp., Integrated Circuit Systems Inc., IC Works Inc., International Microcircuits Inc., NEC Corp., Pericom Semiconductor Corp., Philips Semiconductors, and Texas Instruments Inc. have brought clock ICs to market.
Most major independent memory-module manufacturers have kept pace, offering samples of 64- and 128-Mbyte Direct RIMMs. But the industry-which was expected to begin high-volume production in the first half of 1999-is far from doing so.
"There's not enough product available for Rambus to have gone through a true ramp-up, so we don't even know all the problems we'll have yet," noted Semico's Garber, reiterating her earlier prediction, now confirmed by many vendors, that the technology won't be in volume production until late 1999.
"I'm not really sure that by the year 2000 it's reasonable to expect a large volume of systems based on Rambus," Camintonn's Popovicv said.
Module suppliers admit that the design constraints implicit in the 800-MHz memory-bus architecture are formidable. "The PC board is a big challenge," Smart Modular's Johnston said. "Their tolerances are tighter, and their vias are smaller. So, as the volume picks up, the PCB guys are going to have to get a little bit better at what they're doing in general."
X-ray equipment is required to check that soldering is done properly, said Shannon Biggs, executive vice president of manufacturing and engineering at module maker Viking Components Inc., Rancho Santa Margarita, Calif. "You have to have good system solder joints and be able to verify that statistically through your process."
The lack of reliability data is also an issue, noted John Yealland, director of process development at Celestica Inc., Toronto. The module manufacturer, which is currently sampling RIMMs in 64- and 128-Mbyte densities, has been working with the Jet Propulsion Laboratory and others to test the reliability of different packages, including BGAs and chip-scale packaging.
Differentiating factor
Perhaps the greatest challenge RIMMs present is in the test process. Having the right equipment and using it properly will be critical.
"Today's standard memory-bus speed-100 MHz-really just came into high volume this year," said Dan Simpkins, market analyst for the VLSI test division at test-equipment manufacturer Teradyne Inc., Boston. "Now, system makers are saying they want to move to a new standard at 800 MHz. That's a rapid increase."
Customers are just beginning to define their test processes, said Don McCord, marketing vice president at Tanisys Technology Inc., Austin, Texas. "They have a basic plan as to how they're going to test their Rambus modules, but they don't have enough data at this point to know how to optimize it," he said.
Leading test-equipment vendors, including Teradyne, Hewlett-Packard Co., and Advantest Corp., recently brought out advanced memory-test systems for the Rambus market. In October, for example, Hewlett-Packard announced the HP 95000 memory tester designed to support 1-GHz data rates for Direct Rambus production environments.
But the systems currently available don't come cheap. "Testers today are effectively running about $3 million and up, and that's just for the hardware," McCord said.
"I've been talking to all the module manufacturers recently, and they'll have to spend at least $2 million to retool for Rambus," said Pierre-Yves Lesaicherre, international marketing manager at Philips Semiconductors, Eindhoven, Netherlands. "That's going to be difficult to do, because the last few years have not been very profitable for anyone in the business."
Lead time is another issue. "If you made a decision today to be in the Rambus memory-module business, it would take you many months from when you place your order to when you received your tester," McCord said. "So, if a customer came to a memory-module manufacturer and said, 'We'd like to give you a really big piece of business, can you handle it?' they'd have to wait for 20 weeks or more before they had the tester capacity in place."
Tanisys has a unique perspective on the memory-module test business. Born of a merger between memory supplier 1st Tech Corp. and memory-test system supplier DarkHorse Systems Inc., Tanisys focuses on build-to-order manufacturing of modules and memory-test equipment. While more than 80% of the company's business comes from supplying memory modules primarily to major PC makers, the rest comes from the sale of the company's DarkHorse Sigma-3 memory testers.
Used by a number of major memory- module suppliers to test 100-MHz SDRAMs, Tanisys' Sigma-3 line is designed to attack the high cost of testing. Prices begin at about $100,000, in contrast to the $1 million to $2.6 million other test-equipment vendors charge, McCord said.
Tanisys is reportedly working on an upgrade to the Sigma-3 for RIMMs that would retain the same type of price differential. Moreover, lead time for the low-cost testers is as short as four weeks.
"Customers that have run correlations between our testers and more costly systems from Advantest and Hewlett-Packard say the Sigma-3 catches well over 90% of the defects caught by the more expensive testers at less than 25% of the cost after factoring in test throughput," McCord said. "And if your single high-speed tester breaks, you're out of business; but if you run multiple Sigma-3 testers and one fails, you're just losing a small percentage of your capacity."
Niche or not
While Rambus entails a number of revolutionary changes in module design, new DDR DRAMs offer a less traumatic transition, according to module manufacturers.
DDR SDRAMs boost performance by transferring data on both edges of the applied clock so that a DDR SDRAM clocked at 133 MHz will offer a burst rate of 266 MHz or a peak bandwidth of more than 2.1 Gbytes/s in an 8-byte application. Yet the technology allows memory suppliers to employ the same packaging and test equipment used with single-data-rate SDRAMs.
"It's a very evolutionary technology that will allow us to leverage much of our current equipment," noted Peter Tomaiuolo, manager of the memory product development group at Celestica. "Some of our customers have begun to request DDR products, and we will have a DDR menu that will be in place later this quarter."
Most industry analysts early on have viewed the DDR market as a niche, particularly given Intel's support of Rambus as the primary main-memory technology for future PC designs.
Analyst George Iwanyc at San Jose-based Dataquest Inc. expects DDR to represent only about 10% of the market in two years.
But momentum seems to be snowballing behind DDR, as the prospect of a relatively easy-to-implement, higher-performance memory technology looks increasingly alluring, especially to high-end workstation and server manufacturers. In the past few months, IC vendors have begun to bring product to market. Most are sampling 64-Mbit DDR devices, and a few have announced the first 128-Mbit and, in one case, 256-Mbit DDR SDRAMs.
"We believe it could be a fairly big market, especially with the high-end guys embracing it simply because of its simplicity of application," Kingston's Sutherland said.
Even the key obstacle to wider deployment of DDR devices-the absence of chipset support from Intel-seems to be fading. Many major companies are developing their own chipsets for high-end workstations and servers. And a number of other chipset suppliers say they are working on a low-end DDR solution.
"If you look at it from a total point of view, the infrastructure is quickly coming together," said Mian Quddus, strategic marketing manager at Samsung Semiconductor Inc., San Jose, and chairman of the module committee at JEDEC.
Quddus has led a JEDEC effort to approve a 184-pin DIMM for DDR SDRAMs.
"There are two keys on the existing 168-pin SDRAM module, and basically the new module replaces one of the keys with the additional pins, so the rest of the 184-pin module is the same as the existing 5.25-in.-long single-data-rate device," he explained.
JEDEC is also in the final stages of approving a 200-pin small-outline DIMM with a 0.65-mm lead pitch for DDR SDRAMs that will serve the portable market. A second 200-pin module with a 1-mm lead pitch is also under consideration for systems running above 150 MHz, as well as a third 200-pin module with a 1.27-mm pitch for high-reliability applications.
Most of the support circuitry for the development of buffered or registered modules for high-end applications has also recently come to market.
"Registered modules typically need several additional circuits on the module, including a PLL that does phase alignment, and a couple of drivers that drive the control and address signals from the memory controller on the motherboard into the DRAM," said Philips Semiconductors' Lesaicherre.
Last month, the company announced samples of the SSTL 16857, a 14-bit SSTL_2 registered driver with differential clocks. Since a typical DDR memory module will require between 23 and 27 registered control and address lines, it will need two SSTL 16857 ICs on each memory-module card.
Texas Instruments Inc., Dallas, is also offering registers for DDR modules.
To synchronize signals on individual memory chips, Philips also brought to market the PCK857 PLL. The 150-MHz differential 1:10 SDRAM clock driver is packaged in a 48-pin TSSOP. The PLL is also being produced by Hitachi Ltd., Integrated Device Technology Inc., and Pericom.
Build-to-order reigns
The memory-module market continues to turn to a build-to-order model. "It will be the future of the business because you take the inventory and time out of the supply chain, and there's more profit and less risk," Smart Modular's Johnston said.
PC OEMs are looking for ways to shorten "the part of the cycle when [they] own and have liability for the inventory," said John Sloan, vice president of Celestica's memory business unit.
Celestica is working with suppliers and customers to shorten manufacturing cycle times, so it can build and ship the final configuration that customers want at the latest possible moment. The company has cut to less than a day the time from release of material to the floor to shipment of a finished module.
Other module manufacturers serving the OEM market are pursuing similar strategies.
"Given that Compaq, IBM, and HP say they are all moving toward it, and companies like Dell and Gateway are already there, build-to-order is not only here to stay, but will grow in a significant way," Tanisys' McCord said. "They'll all be early Rambus customers, so it looks like build-to-order and Rambus will be inexorably linked."
-John H. Mayer is a freelance writer based in Belmont, Mass. |
