Rambus to double speed, attack comms market
By Will Wade
EE Times
(12/10/99, 6:41 p.m. EDT)
MOUNTAIN VIEW, Calif. — Rambus Inc., whose latest memory chips have barely begun to crack the desktop PC market, last week disclosed an ambitious road map for its high-speed architecture, including doubling the clock speed and expanding into non-PC markets that are likely to include communications.
The RDRAM has been slow to ramp into PC designs because of yield problems among Rambus DRAM licensees and delays in the release of Intel Corp.'s 820 chip set, designed to interface with Rambus memory. In addition, some memory vendors are hedging their bets by shifting resources to the double-data-rate (DDR) DRAM architecture, which offers faster throughput than existing synchronous DRAM devices but has similar design constraints.
"We plan to double the data rate of our fastest chips from 800 MHz per pin to 1.6 GHz," said Geoff Tate, Rambus' chief executive officer. "We already have the faster technology working in a lab setting as an engineering demo, and we expect to see product announcements by the second half of next year."
Those first products could come from Rambus' DRAM partners, or from nonmemory companies working to implement Rambus technology in new fields. To observers and licensees, the strategy is easier said than done.
"The problem isn't bandwidth, it's cost," said Jeff Mailloux, DRAM marketing manager for Micron Technology Inc. (Boise, Idaho). "Doubling the bandwidth next year is absolutely not going to solve those issues, and it won't help the chips become more popular for PC main memory."
Implementing a new memory design is never as easy as switching masks, Mailloux said, adding that he was under no illusions that adopting a new RDRAM architecture would be anything less than challenging. Micron is building both Rambus and DDR parts.
"Obviously, we'd like to see the first generation get a little more established and stable, and see the costs come down a bit as well, before we run off and do the second generation," he said. "But if that's what our customers are asking for, then we will make our best efforts to provide it."
Besides increasing the speed of individual chips, the com-pany is tweaking the design of the Rambus in-line memory modules used to implement RDRAMs in PCs. Tate said the modules' total bandwidth will increase fourfold once the speedier devices are used next year. That means a single RIMM will deliver up to 6.4 Gbytes/second of total memory bandwidth.
"Increasing the speed of the RIMM is key for the PC market, while increasing the data rate on the individual devices is important for all our target markets, especially the non-PC applications," said Tate. A wider bus?
Intriguing idea with few details
The company has not disclosed exactly how it plans to improve its technology. The most likely scenario, said Steve Cullen, memory analyst for Cahners In-Stat, is for Rambus to modify its interface architecture to wring out more performance.
Cullen also predicted that the faster memory modules might use a wider data bus. Current RIMM designs use a 16-bit bus, much narrower than the buses implemented in SDRAM-based systems. Using a wider data bus would be an easy way to boost RIMM performance, Cullen said.
While the biggest design wins for Rambus to date have been in the PC and consumer gaming markets, the company clearly has its eye on other sectors. High on the list is communications, where Tate said the high-speed memory is already seeing wider use in advanced networking applications. "We are doing very well in those markets, and several companies will be announcing product early next year," he said.
Rambus first announced its intention to crack the communications market back in February 1998. But at that time, industry interest was lukewarm, even in high-end designs. For example, terabit-router makers Avici Systems Inc. and Pluris Inc. have yet to use Rambus technology in their systems.
For Pluris (Cupertino, Calif.), the decision was a matter of risk vs. reward, said Mohammed Rajabzadeh, director of engineering. Because memory bandwidth wasn't a critical issue, Pluris opted not to take a chance on switching to a new architecture. "When you're designing an ASIC, putting [in] the Rambus core adds some risk to your design," he said.
Rambus does get considered every time Pluris starts a new design, Rajabzadeh said. And he found the promised 6.4-Gbyte per-module speed noteworthy. But memory bandwidth isn't enough of an issue to force Pluris and other vendors to make the switch.
"Traditional networking vendors use off-the-shelf DRAMs and SRAMs because they don't need the bandwidth [of Rambus]," Rajabzadeh said. "So far, it's enough."
Aside from going after new market sectors for its memories, Rambus is also looking to create new applications beyond memory. "Our technology is basically a means to connect different chips, so we can use it for areas besides DRAM," said Tate.
Although Rambus is already working to modify its core technology for a nonmemory application, Tate would not disclose which market it is targeting. "We think our chip-connection technology can be used to address a lot of system bottlenecks, and we will be making more announcements about this next year," he said.
Analyst Cullen agreed, saying, "At its heart, Rambus is a high-speed signaling interface. Nobody is in the DRAM business because they want to be; people are in DRAM because they have to be, and as soon as they can find something else to do, they should do it."
Asked to speculate on what other areas Rambus might tackle, Pluris' Rajabzadeh said that I/O processing could be a promising target. "If they come up with some magic trick to boost I/O speed, that would be a plus," he said. "Their best bet is to incorporate that in an ASIC technology."
Rock fight
Future prospects aside, for Rambus and other DRAM makers the present is marked by the rock fight among memory technologies. In recent weeks, the double-data-rate scheme appears to have picked up steam, even as companies like Samsung crank up the pace of their RDRAM production. Samsung announced last week it would double Rambus memory production, to 2 million units a month by February.
Hyundai Microelectronics said last week it has completed all design work for an RDRAM product.
"We've finally finished development of our Rambus chips, and we'd like to avoid having to go through that again," said product marketing manager David Derrough. However, the company is not manufacturing anything.
Citing a lack of demand from PC customers, Hyundai instead has turned on the spigot for DDR parts. It is shipping its 64-Mbit DDR DRAMs in volume and sampling its 128-Mbit versions. Memory rival Micron Technology also has a 64-Mbit DDR chip in the market, and will sample its 128-Mbit device next quarter.
"All of our top OEM customers are telling us that they expect to use a lot of DDR memory next year," said Derrough. Hyundai has confirmed that Compaq Computer Corp. and Dell Computer Corp. are both using Hyundai's DDR chips in upcoming server designs.
It may not be entirely a customer-demand problem that is prompting the Hyundai switch, however. "Hyundai is a little behind the eight ball" on RDRAM yields and die sizes, which may account for its new focus on DDR technology, said Jim Handy, memory analyst for Dataquest Inc. (San Jose, Calif.).
The key advantages of DDR over RDRAM in the server market include lower costs and wider buses linking the memory to the rest of the system. Where Rambus chips garner significant premiums over standard SDRAM, Derrough said there is no premium for DDR, which uses nearly the same manufacturing process as SDRAM and carries only a tiny die-size penalty.
DDR chips can be implemented in systems with bus widths of 64 bits, identical to those now seen for standard SDRAM-based systems, according to In-Stat's Cullen. In contrast, the widest bus for RDRAMs now is 16 bits, one-quarter the width of a 64-bit bus.
Data moves on and off RDRAM chips at up to 800 MHz, four times faster than in a 200-MHz DDR design. However, on account of the bus widths, an RDRAM would have less bandwidth than a 266-MHz DDR DRAM device. While Micron has not abandoned RDRAM for DDR, Mailloux also sees the advantage of the double-data-rate parts. "We think there will be different memory architectures for different markets," Mailloux said.
Meanwhile, Samsung declared last week that it is doubling RDRAM output between now and February to meet customer demand. Samsung had stopped ramping production of Rambus memories in October after Intel delayed its 820 chip set, which was to support the use of Rambus memories. But the company resumed Rambus production last month.
"Many customers came to us because not many companies can supply (Rambus memories)," said Yoon-Woo Lee, president and chief executive officer of Samsung's semiconductor unit. Although Samsung is building about a million 128/144-Mbit Rambus chips per month, there is still a big market shortage of the devices, he said. Hence the ramp to 2 million per month.
—Additional reporting by Robert Ristelhueber and Craig Matsumoto. |
