Global Sources
EE Times-Asia
Stay in touch with EE Times Asia
EE Times-Asia > Memory/Storage

Elpida, Hynix entries further muddy network DRAM waters

Posted: 23 Apr 2002 ?? ?Print Version ?Bookmark and Share

Keywords:Elpida Memory? Hynix Semiconductor? memory? DRAM? network chip?

Network systems designers beware: If the low-latency memory market seemed fragmented before, it is about to be diagnosed as downright schizophrenic. At the end of this month, Elpida Memory Inc. and Hynix Semiconductors will detail similar, proprietary designs that will match up against two separate standards being pushed by competing memory manufacturers.

During a Denali Software Inc. conference in Boston on April 30, the Japanese and Korean companies, which have so far kept a low profile on low-latency DRAM, will outline network-optimized chips based on the forthcoming DDR-II standard. That will make a total of three standards, backed by seven companies?all vying for a market segment that will only account for 20 percent of the total DRAM market by 2005.

In raw dollar terms, however, it could be worth more than $500 million and yield fatter margins than PC main memory.

Elpida and Hynix will run up against Infineon Technologies AG and Micron Technology Inc., both of which back Infineon's Reduced Latency DRAM, and a partnership between Toshiba Corp. and Fujitsu Ltd. that supports the latter's Fast Cycle RAM (FCRAM). Both architectures are already on the market. Samsung Electronics Co. Ltd. is keeping its plans quiet, despite its having licensed FCRAM and listed engineering samples of its "FCRAM compatible" network DRAM on its Web site.

Elpida will propose two new concepts at the Denali conference. The first, tentatively dubbed xDDR-II, will compete with the current crop of low-latency DRAMs and weigh in with a similar cycle time of 25 to 30 nanoseconds.

The second, more ambitious concept will be a next-generation part with a cycle time of less than 10 nanoseconds, closer to an SRAM device but with the density benefits, I/O and pin count of conventional DRAM. Called Ultra High Speed DRAM, it will be a better match to future, 40-Gbit/second networks, the company said.

Additionally, Elpida will market Rambus DRAM devices for networking, as well as high-speed DDR-II chips.

The Hynix game plan is closely aligned with that of Elpida, which Hynix contacted last November to discuss potential specifications. After meeting with a handful of network processor makers, among them Intel, AMCC and Vitesse, Hynix decided it would also support a DDR-II version "with some tweaking of the architecture to reduce the latency," said Joo Sun Choi, director of strategic marketing and applications at Hynix. In short, designers will increase the number of banks to cut down on conflicts and shorten the access routing.

Choi wasn't completely sure when the parts will sample. Hynix hopes to sample a 4-Mbit x 32 DDR part in the third quarter that is nearly twice as fast as conventional DRAM. That will be followed by "netDDR-II," in a 16-Mbit x 16 configuration at 25 to 30 nanoseconds, in the first quarter of 2003 and by an 8-Mbit x 32 part in the third quarter of next year.

With such similar concepts, Choi suggested, it is likely Elpida and Hynix will join forces to promote their standard. They will stress its similarity to DDR-II to persuade network processor and system designers of its easy implementation.

In part, they will note that its standard I/O will enable it to easily switch in or out with lower-performance, commodity DDR-II to address different application demands within the network infrastructure.

Selection bottleneck

Despite the rush to get this first generation of chips onto the market, the early feedback from NPU designers indicates that cycle times are still lagging. That means few decisions have been made over which part to design in for the long product cycles associated with networking gear. "Right now, I think it's a selection bottleneck, not only for the users but also at the DRAM vendors," said Katsuyuki Sato, deputy general manager of technical marketing at Elpida.

The competing standards on the market will also make it difficult for designers, who want assurances that the part they pick won't melt into obscurity because of sourcing problems. Even though the relatively smaller volumes of the network market mean companies shouldn't need the same web of redundant vendors that a PC maker might want, sourcing remains an issue to be reckoned with.

Consider FCRAM. "Toshiba is out of DRAM except for this one part, and Fujitsu is lying low, so the support of Samsung could be crucial," said Lane Mason, a memory analyst at Denali, which supplies memory models and controller designs, primarily to networking customers.

Mason has followed the fragmentation of the market closely enough to know that it's certainly not your everyday DRAM business. "It's not an easy task to define the part, get the necessary second sources so the user will design it in, and have the patience to ramp up and not get a payback for two or three years," he said. "Meanwhile, everybody else is looking at your data sheet and saying they can do better than that, making it a one- or two-generation part that has a lot of resources invested in it."

With so many horses in the race, what are the odds for Rambus? Don't count it out yet, says Los Altos, Calif.-based Rambus Inc. Recent decisions by Intel Corp. and Vitesse to have their network processors support Rambus DRAM add some weight to the company's claim, but some heavyweight network equipment suppliers would still have to chime in to solidify the architecture's position. Rambus is hinting that surprises are in store for the doubters?of which there are plenty.

Rambus battle

The two biggest technical criticisms of Rambus have centered on latency and random access?areas that have provided competitors ample ammunition in their bid to deep-six the memory's prospects. Frank Fox, a vice president at Rambus, has countered those volleys "again and again" by asserting that the Rambus architecture's 32-bank design helps it power through its shortcomings.

"There are two ways to deal with network traffic: One, you can have very short latency so you can respond very quickly, or two, you can basically pipeline by having a lot of different banks. By having a lot of banks, you don't run into conflicts," he said.

Another important factor that will aid Rambus, Fox said, is its commodity status relative to the newcomers. "If it's a highly specialized part that has a large amount of engineering costs associated with developing it, you're recouping that over a smaller volume of parts," he said of the competing standards. The message that commodity parts will look better because they are cheaper is not the usual argument delivered by Rambus.

Elpida and Hynix may have shed some light on their intentions, but the market is looking more muddled than ever. Samsung will be the company to watch, to see whether it forgoes FCRAM and comes out with DDR-II-based designs that more closely aligns with Elpida's and Hynix's strategies.

Analysts have suggested as much. Indeed, Samsung's nearly nonexistent promotion of its network DRAM hints at that direction.

For a little more intrigue, Elpida said that just because it has conceptualized a network memory doesn't mean its decision is set in stone. "We haven't ruled out licensing FCRAM yet," said Jim Sogas, Elpida's vice president of sales.

Said Denali's Mason: "It's going to take a long time to settle down. There are a lot of irons in the fire."

That may be the understatement of the year.

? Mike Clendenin

EE Times

Article Comments - Elpida, Hynix entries further muddy ...
*? You can enter [0] more charecters.
*Verify code:


Visit Asia Webinars to learn about the latest in technology and get practical design tips.

Back to Top