Since its highly successful IPO drew wide attention two years ago, Rambus has faced threats from competing memory systems, memory manufacturers unhappy with Rambus costs, and the explosion of low-cost PCs. But the Rambus rollout in the last three months of the year will be bolstered by two major players: Dell and Compaq.
Warburg Dillon Read analyst Seth Dickson said Dell will snap up two-thirds of the Rambus memory chips in the last quarter of this year and the first quarter of 2000. Rambus first will arrive in high-end servers and workstations.
Another big deal is with Sony, whose next-generation PlayStation will use Rambus memory, Rambus chief executive Geoff Tate said.
The stakes in the battle are huge. The debate over the future of memory technology is central to the long-term plans of world's biggest computer and memory makers--and the price of the Rambus chips hangs in the balance.
The PC133 connection
The most recent Rambus ruckus has come with PC133, a 133-MHz memory technology that runs somewhat faster than the prevailing PC100. Dickson attributes the emergence of PC133 as the factor that caused Rambus's stock slip from its peak above 100 January to its current value in the mid-80s.
But Tate dismisses PC133 and says it won't derail Rambus' Intel-backed plan to become the next memory standard.
"It doesn't affect us at all. It offers little or no extra performance" over PC100, Tate said in an interview this week. "In three or four years, Rambus will be in all PCs."
Dickson agreed that Rambus is a major step ahead of PC100 and PC133. His company's analysis shows that, when actually implemented, PC133 and PC100 will be able to send data at the speed of about 560 megabytes per second, well below Rambus's 1,500.
Intel, Rambus's biggest ally, naturally agrees with Tate that PC133 isn't worth bothering with. Intel is a key player because it sets so many standards for PC designs and itself makes a key component, the chipset that enables a computer's CPU to talk to its memory.
Although there have been some reports that Intel has embraced PC133, the company denies this. "We're going from pc100 to Direct RDRAM," spokesman Dan Francisco said.
Some believed that an upcoming chipset from Intel, the 810-E, would support PC133. Intel declined to comment on the issue, but sources said the confusion probably stemmed from the fact that the 810-E will talk to the processor at a speed of 133 MHz but work with the memory only at 100 MHz.
Intel's next big chipset, called the 820 and code-named Camino, will be the first to support Rambus memory. That chipset is due around September, a delay from the earlier expected summertime launch.
But there are other barriers to adoption besides PC133, chiefly the expense of using Rambus. Not only do memory manufacturers and computer makers have to license the technology from Rambus and pay the company royalties, but the Rambus chips themselves are more expensive.
Three major memory manufacturers, speaking at a Warburg Dillon Read conference Wednesday, all said Rambus is much more expensive to manufacture than the prevailing SDRAM chips used today in the PC100 architecture.
Comparing Rambus and SDRAM 128-megabit memory chips, Micron has experienced a 20 to 25 percent "die-size penalty," said Jeff Mailloux, DRAM marketing manager. In other words, his company can't make as many chips from a wafer of silicon.
Avo Kanadjian, senior vice president of marketing at Samsung Semiconductors, said his company also expects a 20 to 25 die-size penalty with Rambus.
Although Hyundai Electronics currently has only a 10 percent die-size penalty, Farhad Tabrizi, director of strategic marketing, said Rambus' high speed is another problem. "At this high frequency, you have very small margins," he said.
The faster chips have to run, the harder it is for manufacturers to make them. Current memory technologies have been running at 66 and 100 MHz, and manufacturers "all the sudden have to give it 800 MHz," Tabrizi said.
In fact, some companies have been discussing a clock speed of only 600 MHz, said Merrill Lynch analyst Joseph Osha.
New testing equipment adds to the cost as well. Current equipment can check 16 chips at a time, Dickson said. "The throughput is not there," he said.
But Nicholas Konidaris, CEO of chip testing manufacturer Advantest, said prices should be coming down with the arrival in the next two years of 64-chip Rambus testers, which are expected to cost about $6.4 million each.
That may sound like a lot of money, but it's about half as expensive testing the same number of chips with eight of the company's current Rambus machines, which can test only eight chips at a time. Tate acknowledges the extra expense of Rambus but argues that the performance gain outweighs that cost.
Another barrier to Rambus adoption is control issue that memory manufacturers have with licensing someone else's technology, said Osha. "NEC doesn't like the idea of forking over royalties to some startup in the U.S.," Osha said.
That resistance is one of the reasons memory makers have been pushing competing memory systems, including PC133, PC166, double-data rate DRAM, and SyncLink DRAM, Dickson and Osha said.
Do cheap PCs undermine Rambus?
Another factor that's changing the memory landscape is the emergence of low-cost PCs, machines that may not have the fastest performance but that are good enough for huge numbers of customers.
A parallel exists here with DVD drives. The new drives are replacing existing CD-ROM technology more slowly than expected earlier because most ultra-cheap PCs still come with CD-ROMs, analysts have said.
"We agree that there's a bifurcation at the low end of the market," Tate said. But he added that people always will need faster PCs. "What's good enough today is ten times faster than the rocket science ten years ago," Tate said.
Moreover, "The high-performance portion of the desktop market is 20 percent. That's pretty high volume," Tate said.
Osha said Rambus' value for lower-end systems is tied to fancy features such as multimedia and streaming video.
Micron's Mailloux said there's room for several types of memory. "We don't really believe that a one-DRAM-fits-all approach is the path to the future," he said. "The market is segmenting. There's no reason a PDA [personal digital assistant] will use the same RAM as a high-end computer."