Rambus faces a rocky road

Although the company has been virtually anointed as the standard bearer for the future of computer memory, Rambus has been saddled with product delays and technical glitches in 1999, which in turn have delayed the debut of Rambus-based PCs from the middle of the year to the end of the third quarter. The company yesterday reported single digit growth in revenues and earnings for its second fiscal quarter.

But even more critical, the company could start to feel the squeeze of cheaper alternatives and the industry's obsession with keeping desktop prices as low as possible. The technology is great, nearly everyone agrees, but how many people are going to pay for it?

"You've got to pay a substantial premium with Rambus, so that puts the burden on Intel and Rambus to demonstrate that the performance benefits are there and that they aren't available for the other platforms," said Nathan Brookwood, a consultant with Insight 64. If the companies can make their case, the debate ends, he added. But for now: "It's cloudy."

Others strongly discount the notion, stating that technical glitches and revenue streams won't matter until the second half of 2000, when Rambus is expected to be more of a volume product. Still, some risk exists. "There is an 80 percent chance that Intel sees this thing through," said Drew Peck, an analyst at Cowen & Company. "But if Intel were ever to waver, the whole thing is shot."

While the debate may sound arcane, it has significant implications for the industry. Intel has invested $600 million into memory companies to ensure that supplies of Rambus memory will be ready for sale this fall. Without Rambus-style memory, PC makers will not likely be able to wring all of the performance out of the 600-MHz Pentium III.

Intel has developed "work-arounds" so that PC makers can graduate to faster Pentium IIIs with current memory chips. Alternative solutions from third parties will do the same. The question, therefore, is whether the substitutes eventually will become more attractive.

"They can't hold back the PC industry for Rambus," said Ashok Kumar, semiconductor analyst at US Bancorp Piper Jaffray. Using Rambus in a PC instead of standard SDRAM (synchronous dynamic RAM) will initially add around $100.

"It is not enough of a performance differentiator. The price is 50 percent more per megabit. The question is who is going to absorb it," Jaffray said.

A little history
The Rambus memory system helps to ameliorate the growing speed disparity between computer CPUs and memory. As CPUs have gotten faster, it's been harder and harder for a computer's memory to keep up. Rambus closes the gap by accelerating the speed at which data is transferred. Currently, the peak bandwidth for Rambus memory s 1.6 gigabytes per second, twice the peak bandwidth of current SDRAM.

The company enjoys a huge following. All of the major memory manufacturers have licensed Rambus' memory architecture and have announced that they will produce products. Seven of the top memory makers, in fact, have already started releasing RDRAM, the name given to Rambus style memory.

The company's watershed moment was undoubtedly its alliance with Intel. In late 1996, Intel announced that it would support Rambus in its future chipsets. That alliance made Rambus the de facto standard for future memory, according to many. National Semiconductor, AMD, and Compaq, which controls the Alpha processor, have all taken out Rambus licenses.

While the future for the company has largely been rosy, 1999 hasn't been its best year. In February, Intel delayed its 820 chipset, code-named "Camino," from June to around September. This effectively delayed the release of PCs using Rambus until the first part of the fourth quarter, because the 820 is Intel's first chipset that can support Rambus.

Another feature of the 820, however, is the fact that it is Intel's only chipset with a system bus that runs at 133-MHz, faster than the 100-MHz system bus that currently comes with Intel's chipset. This is crucial because 100-MHz chipsets will not work with Pentium IIIs running at 600-MHz or more, according to many. In other words, if delays to Rambus memory lead to delays in the 820 chipset, PC makers can't pick up the 600-MHz Pentium III.

Intel, of course, has not left this up to chance. Along with the 820, Intel will release the MTH, or "Memory Translator Hub," according to sources. The MTH is a converter chip that will allow PC makers to combine standard 100-MHz SDRAM, the 820 chipset, and faster Pentium IIIs. The only problem with the converter chip, however, is that it detracts slightly from performance, said Dean McCarron, principal at Mercury Research.

Customers will likely use Rambus as well as standard memory with a converter, said Paul Otellini, general manager of the Intel architecture business group. "I wouldn't categorize it [Rambus memory] as a requirement. The 133-MHz bus can scale independently of Rambus."

Sources close to Intel, however, state that there are work-around products in early 2000 that will follow the first MTH, which raises the possibility that a long conversion could start to make Rambus less attractive.

Low-cost alternative
More significantly, alternative chipset makers are short-circuiting Rambus in a different way. Via Technologies is releasing a chipset with a 133-MHz system bus. Memory makers, moreover, are finding it fairly easy to bump their standard DRAM from 100-MHz to 133-MHz, according to Brookwood. In the end, this creates the existence of a low-cost alternative.

"Cost, that's the big question," said McCarron.

Rambus executives agree that memory based on their designs cost more, but not much. The premium comes to only around 15 percent, and will go down over time. Others disagree. Kumar, Brookwood, and others have presented a range that extends from 10 percent to 70 percent.

Technically, Rambus hasn't done itself any favors this year either. Although earlier the first Rambus memory chips were expected to run at 800-MHz, the company now says that memory makers will release 600-MHz and 700-MHz RDRAM chips.

This is an indication that memory companies aren't able to produce enough of the faster chips, said Danny Lam, an analyst with Fisher-Holstein.