Rambus: Friend or Foe?
Its revolutionary technology and marketing savvy have put this small chip company on the map—and offended some of the biggest names in the business
Julius Caesar might have sympathized. On 15 March, the proverbial Ides, shareholders of the semiconductor design firm Rambus Inc. looked on in shock as their stock went into free fall, losing nearly half its value in a single day. The cause: a news report hinting that the Los Altos, Calif., firm's looming patent infringement suit against Infineon Technologies AG, of Munich, had hit a snag. Rumor had it that the judge was about to accept Infineon's definitions of certain technical terms in the four patents under dispute. What sent the stock southward, though, was not so much the specifics, but the growing sense that Rambus' entire legal case was unraveling.
The case is just the first of seven pitting Rambus against Infineon and two of the other biggest chip makers in the world, Micron Technologies and Hyundai Electronics. Since its founding 11 years ago, Rambus has managed to leverage its high-speed memory interface technology into something approaching an industry standard for the dynamic RAM, the most widely used type of memory. The company claims that the patents for Rambus DRAM, or RDRAM, also cover aspects of two competing memory types, namely, synchronous DRAM and double-data-rate synchronous DRAM (known as SDRAM and DDR, respectively).
These memory types, though, were intended as open standards, hammered out by committee under the aegis of the Jedec Solid State Technology Association, in Arlington, Va. Rambus' insistence that chip makers pay royalties for SDRAM and DDR has earned it the wrath of many in the semiconductor industry. Although seven memory chip makers have agreed to Rambus' terms, all eyes are on the current court battle.
The stakes are high. Should Rambus lose, its current licenses for SDRAM and DDR will undoubtedly collapse, along with the leverage it sought to gain for its own memory technology. Infineon, Micron, and Hyundai stand to lose much more, said Julia Elvedge of Chipworks, Inc., a semiconductor analysis company in Ottawa, Canada. "If they lose, they will need to reach a settlement with Rambus, or face an almost total shut down of their DRAM business."
In the end, though, these cases are just the legalistic manifestations of a much broader, industry-wide rift—if not an all-out war—pitting Rambus' supporters, most notably Intel, Samsung, Toshiba, and Dell, against the likes of Advanced Micro Devices, assorted chipset makers, and, of course, Micron, Infineon, and Hyundai. At issue is what will become of, and who will control, the technical standard underlying the US $29 billion DRAM market. The industry now stands at a crossroads: to go with Rambus' proprietary technology or choose the more conventional Jedec-defined approach.
"This is really bloody—it's a fight to the death," is how Art Nutter, an expert on semiconductor intellectual property, summed things up. "If Rambus wins, they'll be one wealthy company." The Infineon trial will set a strong precedent for the remaining cases, he added.
As it happened, the rumors of 15 March proved accurate. Infineon's advantage was short-lived, however. Just hours after the decision, a cache of long-lost documents, recovered from an old hard drive at Infineon headquarters, was turned over to Rambus' attorneys. Among the 4000-odd pages was a 1992 memo that proposed making a "public domain" version of the Rambus technology. This, a Rambus lawyer told the judge, proved the defendants had "cherry-picked" from Rambus designs while drafting the synchronous DRAM standard. It also seemed to directly contradict what Infineon executives had said in earlier depositions.
The judge noted that the new material looked like "important stuff" and agreed to postpone the trial. At press time, it was set to begin on 20 April in Richmond, Va.
The need for speed
A visit to the Rambus offices, overlooking a nondescript stretch of El Camino Real, confounds expectations. The company moved here in January, but apparently there's been no rush to get settled: cardboard boxes sit stacked next to the receptionist's desk; a temporary plastic company banner flaps gently outside the entranceway. The whine of a workman's drill cuts in and out, but otherwise, it's very quiet. Can this low-key outfit really be the nerve center of the high-tech upstart that's been tearing up the semiconductor industry?
Well, yes. Even as a start-up, "Rambus was very boring," recalled cofounder Mark Horowitz. It avoided much of the typical volatility of a young company: staff turnover was low, there were no management upheavals, and no one got divorced.
In other ways, though, the Rambus story began like so many Silicon Valley tales. In late 1988, Mike Farmwald, a University of Illinois professor with an entrepreneurial streak, called up Horowitz, an electrical engineering professor at Stanford University, in California, with the latest of his many neat ideas: a way to close the ever-widening speed gap between computer microprocessors and memory chips. [See "A Rambus chronology.]
Starting in the mid-1980s, central processing unit (CPU) speeds had begun to soar, while memory speeds improved only marginally. Such a mismatch is terribly wasteful: it leaves an otherwise fast microprocessor cooling its heels while data gets read from memory.
Successive generations of both processors and memories addressed the problem piecemeal. But as more demanding applications began to emerge—like streaming audio and video, three-dimensional imaging, and data mining, as well as the need to perform many of these tasks simultaneously—such incremental fixes were quickly eclipsed.
Farmwald and Horowitz talked over dinner. "I have a reputation of telling people their ideas won't work, then going off, thinking about them for a day, and coming back and saying, 'Well, you could do it this way,' " Horowitz said. "This time, it took just to the end of dinner." Although the idea seemed feasible, he told IEEE Spectrum, "I didn't see how you could make a business from it."
Farmwald thought otherwise. That summer he quit his job and moved back to California, and he and Horowitz began working through the details.
To achieve a very high rate of data transfer between the memory chip and the memory controller, they replaced the standard 32-bit-wide memory bus—which back then topped out at 33 MHz—with a narrower but speedier bus. This bus, which came to be known as the Rambus channel, would be 9 bits wide (8 bits for data, one for parity), run at a clock rate of 250 MHz, and read or write using both edges of the clock pulse. Systems using Rambus technology would be able to transfer data at 500 MB/s—more than four times as fast as existing DRAM [see "Rambus Memory in Action"].
Investors were immediately won over, most notably Bill Davidow, who became Rambus' chairman of the board and whose conference room, at the Menlo Park headquarters of the venture capital firm Mohr, Davidow Ventures, served as the company's first office.
The memory, logic, and systems companies that Rambus approached were more skeptical. In terms of the protocol, electrical signaling, and components, it was a radical reworking. "Most of the people we talked to said we were crazy, because we were adding complexity to increase bandwidth," Horowitz recalled. The Rambus memory would not only cost more, at least in the short run, but would also take slightly longer to retrieve the first word of data in order to decrease the time to get the rest of the data. "That was controversial at the time," Horowitz said. "But if you look at the high-bandwidth DRAMs now, they're all using things we talked about back in 1990."
None of the chip makers were biting. Privately, though, they were taking Rambus quite seriously. And rightly so, said Joel Karp, a 35-year industry veteran who served as Rambus' vice president of intellectual property from 1997 to last year. "Rambus' RDRAM is a disruptive technology," he noted. Adopting it would entail all kinds of changes in chip testing, packaging, and motherboard design. Indeed, a 1992 memo that made the rounds at Siemens Semiconductor called the technology "a deadly menace to the established computer industry." At least a couple of companies considered buying Rambus outright.
Finally, in 1995, Silicon Graphics Inc., of Mountain View, Calif., agreed to use Rambus in its high-end workstation and later in its graphics system for the Nintendo 64 videogame console. But the real break came in late 1996, when Intel decided to develop Rambus-based chipsets for its Pentium 3 processors.
Fortifying the Rambus designs into a robust PC memory—one that could be reliably produced in volume—proved "horrendous," an engineer familiar with that effort recalled. The Rambus channel was widened from 9 bits to 16, and the clock rate was boosted to 400 MHz. At such high frequencies, though, the signal integrity could not be maintained; the eventual solution was to limit the number of RDRAM modules to two per board.
In the end, the Intel chipset went into production and the first Rambus-based PCs debuted at Comdex in November 1999—just three months behind schedule, but too late to catch much of the holiday and end-of-year buying season.
A different kind of business
In addition to its novel technology, Rambus has also pursued an unusual business model. The company does not make chips, or anything else for that matter. Instead, it licenses its designs to manufacturers, collecting royalties on whatever products emerge. Rambus now holds 104 U.S. patents, 29 of them based on Horowitz and Farmwald's original April 1990 application.
Rambus didn't set out to be an IP company, said Rambus' chief executive officer Geoff Tate. "It was a Catch-22. We didn't have the billion dollars to build a wafer fab, and even if we did, people don't want to buy from just one source, especially a brand new company on the scene." The conclusion, he said, was that "we should go to the DRAM companies and teach them our inventions and ask them to compensate us for the use of our inventions."
To be sure, Rambus was not the first to adopt such a model, even in the semiconductor industry. MIPS Technologies Inc., in Mountain View, Calif., and ARM Ltd., in Cambridge, UK, both predate Rambus. (In fact, Horowitz and Farmwald first met while consulting for MIPS.) But, Tate said, Rambus founders didn't really consider those companies as models. They looked instead to companies like San Jose, Calif.-based Adobe Systems Inc., whose Postscript code became a standard for laser printers; and San Francisco's Dolby Laboratories Inc., which turned its audio technology into a household name.
Much of the game plan, though, had to be worked out as they went along. In the beginning, Farmwald predicted that it would take six engineers about a year to perfect the Rambus interface and get the industry to adopt it. It was a wild under estimation, and became a favorite in-house joke.
"Early on, we thought the chip companies would be more active in marketing our technology," Tate said. "We soon realized that nobody cared about our technology as much as we did." Getting a chip company to sign a license is one thing, but getting it to make Rambus devices is quite another—and getting those devices built into products is yet another. After the first year, Rambus sharpened its sales and marketing efforts.
The strategy has paid off. Having essentially no capital expenditures, the company first became profitable five years ago. Since going public in May 1997 (the same month as another company known for its unique business model, Amazon.com), Rambus' stock has risen as high as $127, after a four-to-one split in June 2000. And, while industry acceptance came slower than Farmwald had predicted, at last count Rambus technology was shipping in over 250 products, including workstations, desktop PCs, game consoles, and high-definition TV sets [see graph]. It has also signed licensing agreements with nearly all of the big memory chip firms. In March, Samsung, Toshiba, and Elpida announced plans to rev up their production of RDRAMs significantly.
"They have an incredibly astute business strategy," noted Terry Ludlow, president of Chipworks. Many of his clients are keenly interested in what Rambus is doing, although he won't say on which side they come down. "It's fairly remarkable that a little company can generate so much passion and wield so much influence," he told Spectrum.
A technical challenger
At least some of that passion has been channeled into developing DDR, an alternative high-bandwidth memory architecture. Like the Rambus memory, a DDR chip uses both clock edges for sending signals. Unlike RDRAM, though, DDR is an open standard developed through Jedec and thus carries no licensing fees. It also builds on existing designs—it's by and large an extra patch of circuitry behind the data pins of a synchronous DRAM chip—so the costs of implementation should be somewhat lower. Among DDR supporters, the rallying cry is "evolution, not revolution."
Mainstream memory is just like any other commodity, said Desi Rhoden, the chairman of Jedec as well as president of Advanced Memory International Inc., a San Jose, Calif.-based industry group that is promoting DDR's adoption. And so there's incredible pressure to keep costs low. "In order to make it dirt cheap, we tend not to use technology on the bleeding edge," Rhoden said. "There's nothing in DRAM that hasn't been around for years." From the customer's perspective, he argued, "adding more [standard] memory is far superior to putting in a different kind of memory."
Former Rambus officer Joel Karp agrees that the marketplace will ultimately decide whether Rambus or DDR or something else prevails, but disagrees on the likely outcome. In the early 1990s, he was Samsung's vice president of strategic product planning and a Rambus skeptic. Then he looked at the technology. "Horowitz and Farmwald basically did what people said couldn't be done," Karp told Spectrum. "I think it will be years before the world realizes just how brilliant they were."
In fact, though, Rambus already enjoys a certain celebrity outside the industry. There are several Internet message boards dedicated to Rambus; a quick scan leaves one with the distinct impression that what's being discussed there is less an engineering company than a sports team.
One diehard fan has even created a Web site, www.rambusite.com, devoted to news about the company. Jeff Fletcher, an accountant in Lonoke, Ark., said he started the site because, as a Rambus shareholder, he grew frustrated at the attacks on the company "coming from certain sectors of the electronic trade press." The site has proved popular, racking up 60 000 hits on a busy day.
Fletcher seems genuinely pained by the controversy surrounding Rambus. "I don't understand the animosity," he said. "What I really wish is for everyone to kiss and make up."
Full court press
It seems unlikely that Jeff Fletcher will get his wish. Rambus' legal battles began in January of last year and show no signs of letting up. First the company sued Hitachi for patent infringement; that case was soon settled out of court. Then, last August, Micron and Hyundai filed suits against Rambus, seeking to overturn its patents. Rambus in turn filed patent infringement suits against Micron, Hyundai, and Infineon.
To be sure, lawsuits over intellectual property are nothing new in the chip industry, and for Rambus, whose patents are its life blood, they may be essential. Writing about the Rambus portfolio in his company's newsletter, Chipworks' Ludlow noted that the patents "benefit from good prior art disclosures and well-targeted claims that [appear to cover] current SDRAM and DDR SDRAM. They will be difficult to attack based on technology only."
But Rambus is also accused of misconduct—specifically, of violating Jedec's open standards rules by incorporating into the Rambus patents technology that Jedec was discussing for SDRAM and DDR. Micron and Hyundai further allege that Rambus sought to gain a monopoly over high-speed memory technology.
Also on trial, indirectly, is Jedec. Its job was to decide what the industry standard for DRAMs will be. But if the group's specifications can be legally, if not technically, trumped by a proprietary architecture, what purpose does Jedec serve?
When asked by Spectrum about the company's ongoing legal disputes, CEO Tate would only say, "We didn't do anything wrong." But during the company's annual meeting in January, he gave a more complete defense: the Rambus nondisclosure agreements signed by the chip companies in the early 1990s prove they knew what Rambus was working on well before its engineers joined Jedec.
The new new thing
Amidst all the legal wranglings, engineers at Rambus and Infineon have continued to work together. "The lawsuits have had no effect on the working relationships between the engineers," noted Frank Fox, Rambus' vice president of engineering. Indeed, both Infineon and Hyundai (recently renamed Hynix Semiconductor, Inc.) just finished validating their 288-Mb Rambus DRAMs, clearing the way to mass production later this year.
Fox's team is also refining a new signaling scheme that promises to double the RDRAM bandwidth to 1600 Mb/s per pin. Called QRSL (for Quad Rambus Signaling Level), it is aimed at consumer, graphics, and networking applications. The QRSL effort started five years ago, and it may be another two years before the technology begins to pay off. "It's a neat concept that we've refined through practice," Fox said. "Test chips and physics keep you honest."
Rambus also has designs on the non-memory market, according to Steve Chen, Rambus' vice president of marketing. That is only logical, he said, since the DRAM share of the total semiconductor market is shrinking. Last year the company introduced its first networking product, a high-speed backplane connector called SerDes.
In the end, it comes down to this: for an IP house like Rambus to survive, figuring out the next big idea must be an all-consuming task. The court contests, the animosity, the media scrutiny—these are just distractions. "We're an innovation company," is how Tate put it. "If you're going to sell ideas, you'd better have innovative stuff. If your ideas are just ho-hum, people are going to say, 'Why pay Rambus for that?'
To Probe Further
Rambus Inc.'s proprietary scheme for speeding up communication between processor and memory was described in the October 1992 issue of IEEE Spectrum. The Rambus Web site, https://www.rambus.com contains information about the company and its technology. Jeff Fletcher's https://www.rambusite.com is also useful, in particular the message boards and legal documents from the company's various court battles. The specification for double-data rate (DDR) synchronous dynamic RAMs may be found at https://www.jedec.org the Web site of the Jedec Solid State Technology Association in Arlington, Va.