Paul:
Would you please translate what all this means to long term INTC investors. I'd appreciate your comments as to what this technology will enable in 1999.
Thanks.
Barry
From CMP's OEM magazine:
<<February 01, 1997, Issue: 536
Section: Features -- Cover Story
Carpe DRAM -- Is Asia's dominance a memory?
By David Lammers
According to Intel Corp.'s road map, by early 1999 the
desktop PC industry will begin the transition to processors
that are fed data at 1.6 Gbytes per second from a 64-Mbit
RDRAM, a memory chip now being developed by a small
Mountain View, Calif., company called Rambus Inc. A 1990
startup, Rambus expects about 15 percent of the Intel-based
desktops shipped in the last year of the decade to use its
RDRAM technology, rising to 35 percent in 2000 and about
65 percent in 2001. That scenario means that memory-chip
makers, once the powerful bellwethers of the semiconductor
industry, must license Rambus technology and start volume
64-Mbit RDRAM production in the final quarter of 1998-not
much time for any companies that have yet to start a Rambus
program. Today, only two DRAM makers, NEC and
Toshiba, have experience making Rambus DRAMs in
quantity, and they have been making only 16-Mbit parts. The
impact of the Intel-Rambus connection on the DRAM
industry promises to be immense, prompting fears that one of
Asia's remaining technology strongholds is about to become
little more than a caboose on the Intel-Rambus locomotive.
More profound consequences may follow. Some observers
expect that at some point, Intel will build a Rambus interface
into its processors, rather than using a Rambus-enabled
memory controller, bypassing the need for core logic in
entry-level PCs. What's more, level-two SRAM cache may
be eliminated in many future Intel systems as just another
obstacle standing between the processor and the ultrafast
RDRAM channel. On top of all that, Microsoft Corp. is
licensing a next-generation graphics architecture, Talisman,
that also calls for RDRAMs.
Nor will those companies designing alternatives to the Intel
PC be immune from the technology transition. Apple
Computer Inc. and partners Motorola and IBM, as well as
X86 clone makers such as Advanced Micro Devices and
Cyrix, must decide how they will meet the
memory-bandwidth challenge.
The tsunami that's about to sweep the industry had its origin
in 1987, when a Stanford PhD named Mike Farmwald came
up with an interesting idea. Working on an ECL-based
processor for use in high-end workstations at MIPS
Technologies Inc., the pioneering RISC processor maker,
Farmwald found that developing a memory interface that
could keep up with the R6000 was tougher than designing the
processor itself. He saw that one key to feeding the processor
quickly was to shorten the physical distance between it and
the memory. The result, which Farmwald calls "a little
backplane bus," sparked Rambus.
Farmwald left MIPS to teach for a year at the University of
Illinois but returned to discuss his ideas with Mark Horowitz,
a fellow PhD at Stanford. "Mark's general philosophy is to
attack a new idea with every ounce of his intellect and, if it
survives, to support it," Farmwald says. "He looked at my
CMOS drivers and said they were garbage but thought he
could make it all work."
Horowitz wanted to turn Farmwald's proposal into a Stanford
research project but, at the end of 1989, Farmwald "was
adamant about forming a company." The pair spent the next
two years working on the idea with a small group at what
became Rambus.
Once the company was incorporated, "there were more
people interested in funding the deal than we really needed,"
says Horowitz, testimony to the venture-capital mentality in
Silicon Valley, where "a couple of guys with a wild-ass idea
could get sane people to actually sit down and listen." The
VCs liked what they heard: a radically new approach to
increasing memory bandwidth. The first convert was William
Davidow, who had recently left his job as Intel senior
marketing vice president to set up a venture firm. He became
the chairman of the Rambus board and continues in that role
today.
Recognizing the potential value of the Farmwald-Horowitz
ideas to a microprocessor vendor, Davidow picked up the
phone and called Intel's chairman Gordon Moore. Moore was
in a meeting. As Farmwald tells it, Moore's secretary asked if
it was urgent, Davidow said yes and the Intel chairman came
to the phone to hear his spiel. At a meeting hastily set up for
the following day, Moore immediately appreciated the
concept the three men put forward.
Indeed, he liked it so much that Intel took a share of the
fledgling company and laid out plans to implement the
Rambus interface in the 486 processor. That proved
premature, given the ability of conventional DRAMs to feed
the 486 and the lack of a PCI bus that could handle graphics
data quickly. A clause in the Intel investment agreement
allowed Rambus to buy back its shares if Intel did not use its
technology in the 486 generation. In order to raise more cash
from venture capitalists, Rambus bought back the shares at
an agreed-upon (low) price and resold them to its VC
investors.
"When we look back on that now, it was a pretty stupid thing
to do," Farmwald says. "At the time, I don't think we ever
got anywhere close to the point where we couldn't meet the
payroll, but we had only a few million dollars and we wanted
to use those shares to raise cash from the venture-capital
backers." As a result, Intel's attitude toward Rambus chilled
for a time.
After getting Moore's ear, Davidow looked to another Intel
colleague: Takahiro (Tom) Kamo, the co-founder of Tokyo
Electron Ltd., an Intel distributor for several years, and the
former president of Intel's Japan subsidiary. In 1991, as
chairman of Rambus' Japanese subsidiary, Kamo took the
Rambus concept to executives at three of Japan's top
semiconductor makers: Hajime Sasaki at NEC, Tsuyoshi
Kawanishi at Toshiba and Matami Yasufuku at Fujitsu.
In a joint press conference in March, 1992, the three
announced themselves as the initial Rambus licensees, though
Yasufuku's successor at Fujitsu allowed its Rambus activity
to lapse. The company is renegotiating a license now that
Intel has set down marching orders for the DRAM industry.
Kamo says that Toshiba's Kawanishi in particular realized
that Rambus, as "an outsider," was in a position to set a new
DRAM standard, something no single big DRAM vendor
could do, given corporate rivalries. In the past five years,
Hitachi, Hyundai, LG Semicon, Oki and Samsung have also
hopped on the Rambus bandwagon, along with a large
number of logic "RAC" (Rambus ASIC cell) licensees.
After this initial burst of activity, Rambus entered an ice age
that might have killed a less persevering company. Though
Toshiba developed Rambus DRAMs at 4-Mbit density, there
were few takers. "For a number of years, in the 4-Mbit
generation, the RDRAMs didn't sell," says Farmwald. "I can
look back and see now that even though the interface
technology was great, the customer had to implement a new
piece of logic on his controllers. We underestimated how hard
that might be. You have to remember that the engineering
resources at a lot of these PC companies are very thin. Even
a big PC assembler usually will have a small staff of
engineers."
Geoff Tate, who once ran the Intel-compatible
microprocessor effort at AMD, joined Rambus in 1990 as
CEO. A Canadian, Tate has an MBA from Harvard to
complement his technical education. More important, his
calm, almost unflappable personality and steady hand on the
tiller are credited with keeping morale up during the years
when Rambus's sales were nearly zero.
Rambus spokeswoman Julie Cates says the early 1990s were
a period of nay-saying: that the DRAM will never work that
fast, that Rambus DRAMs will be as costly as video RAMs,
that high-speed designs will require expensive pc-board
technology, that Intel will never take a license and so on.
Even so, Rambus never went begging. The company has
spent about $50 million since its inception, only about $15
million of it from the VC community. License fees provided
the bulk.
On the plus side, the dearth of real business in the 4-Mbit
generation gave Rambus engineers time to learn about their
licensees' processes, about packaging and about how to
develop test vectors and reference designs. In terms of logic,
Rambus learned how to integrate its interface cell into
graphics controllers at Cirrus Logic Corp. and other early
adopters. "We learned so much in that period when we didn't
have many customers," says Cates. "The DRAM vendors
were so short on capacity anyway that it was hard to get our
licensees to put many engineers on the Rambus project
team."
Gradually, though, results started trickling in. Engineers from
Silicon Graphics, Rambus and Toshiba worked on a graphics
subsystem for an SGI workstation, for example. And Tate
and his colleagues kept going back to Intel, over and over,
trying to stoke the fire at the company that had the power to
catapult Rambus into the big time. But it wasn't any of them
who helped win over Intel; it was Mario.
Nintendo Ltd., the Kyoto-based game maker, was under
siege. Sega Enterprises Ltd. had usurped its near-monopoly in
the videogame market, then Sony Corp. in 1995 delivered its
Playstation machine with a game that awed critics: Ridge
Racer. As the pressure on Nintendo escalated, NEC executive
vice president Sasaki revealed that NEC, SGI and Nintendo
were co-developing a 64-bit game console that would employ
a Rambus channel between the graphics controller and two
RDRAMs. NEC would use 0.35-micron technology, thought
to be too expensive for a game machine, but would keep
costs down by using the relatively simple circuit boards and
low-pin-count packages made possible by the Rambus
architecture.
The result: the Nintendo64 system, with a MIPS processor, a
separate graphics processor with the Rambus interface and
two 18-Mbit RDRAMs. Despite design and production
delays-and with little more software than a 64-bit version of
Nintendo's mainstay Mario Bros. game-Nintendo's new
machine began to sell. Some 4 million were snapped up in
Japan and the United States in the second half of '96,
launching a volume market for NEC and Toshiba's
RDRAMs.
Kamo came up with a parlor game of his own. The Rambus
chips come in thin, plastic, small-outline packages with only
32 pins, and the channel can be implemented in a two-layer
printed-circuit board. Kamo would go out to meet Japanese
customers and the press with three circuit boards: from the
Sony Playstation, the Sega Saturn and the Nintendo64.
"They could see how simple the Nintendo board was," Kamo
says, "and that was more effective than explaining it in
words."
Kamo's briefcase examples impressed Intel's engineers, who
had formed a DRAM working group to investigate the
various approaches to high memory bandwidth. Dennis
Lenehan, the Intel manager in charge of interfacing with
DRAM makers, says the Nintendo64 proved-at a time when
real-world alternatives were lacking-that the Rambus
architecture actually works.
Several announcements, but no silicon, have come from the
Synclink consortium, which aims to push a
synchronous-DRAM (SDRAM) standard as a license-free
alternative to Rambus. Its members include such Rambus
holdouts as Fujitsu, IBM, Micron Technology, Mitsubishi,
Siemens and Texas Instruments. Some DRAM makers want
to raise the speed of SDRAMs by reading data off the rising
and falling edge, but this so-called double-data-rate approach
is unproven technology.
Intel came to the conclusion that SDRAMs have limited
extendibility, Lenehan says. The fastest SDRAM dual in-line
memory modules (DIMMs) can go is 100 MHz, he says, "but
we have this almost insatiable need for more bandwidth. We
can change the IC interface to SSTL [stub-terminated
transceiver logic], but then we still need a different module
than the 168-pin DIMM, and there are some real significant
system costs." More important, says Lenehan, "is that soon
after that, we have to go beyond 100 MHz, and having made
the changes to the DIMM, the interface and the layout-having
gone through that substantial thrash-all that could only take us
through one processor generation." Meanwhile, the Mario-64
game not only ran fast, but it ran on a relatively cheap circuit
board using DRAMs in a low-pin-count package. And it did it
in 1996.
Thus, Intel made the decision to use the Rambus architecture
for PCs starting in 1999. The chip maker's choice, which
came to light last November, underscores the importance of
memory bandwidth in bringing video and 3-D graphics to
desktop PCs. Intel engineers already have started
collaborative work with the technical staff at Rambus to
extend the current 600-Mbit/s Rambus technology so that
graphics and video data can zap across the short-but-speedy
Rambus channel at 1.6 Gbytes/s in those 1999 models.
It is ironic, in a way, that a made-in-Japan game character
persuaded the U.S. chip giant to adopt an architecture that is
now causing angst among Asian DRAM makers. In Japan
and neighboring South Korea, where DRAM manufacturing
skills are a source of national pride, people often say
"America is good at logic; Japan and Korea are good at
memory." Thanks to the Rambus DRAMs, the aphorism just
doesn't ring true anymore.
To some Asians, Intel's decision represents a defeat, a sign
that their own creativity and foresight have been lacking.
Kamo downplays that idea. "Maybe there is an NIH [not
invented here] feeling," he admits, "but the fact is, somebody
has to set the standard." No longer will it be the slow,
democratic committee structure at Jedec (the Joint Electron
Devices Committee), where-as Rambus marketing vice
president Subodh Toprani argues-one maker can object to
another's ideas and delay progress.
What scares Asian manufacturers is that the key DRAM
interface will be controlled by a company on the wrong side
of the Pacific Ocean-an enormous degree of power for little
Rambus. Rambus will continue to manage the licensing of its
intellectual property, and it will be Rambus engineers who
work with the various graphics, chip-set and DRAM makers
seeking to implement the interface in their devices.
Worse, behind Rambus stands Intel-a company that has
rocketed past all semiconductor companies in size, earnings
and raw power-contributing its enormous influence and
knowledge of the PC architecture. To be sure, Intel's
effective control of DRAM interface technology dates back a
few years. Even for the 66-MHz SDRAM, it was Intel that
laid down the law, detailing the various timing specifications
that the memories had to meet to get the stamp of approval
for use in Intel-based systems.
Aside from injured pride and loss of technological control is
the matter of money. With Texas Instruments still tightening
the screws for royalties on its fundamental DRAM and other
patents, along comes Rambus, offering engineering services
and patented technology for a license fee that is partly based
on per-chip royalties. When Rambus was barely surviving
and the DRAM makers were printing money, fees didn't
seem so important. But now that the DRAM makers face
rock-bottom prices for today's products, the prospect of
paying 1 to 2 percent to Rambus is less than appealing.
A senior executive at a Korean DRAM maker was careful in
his choice of words: "I think you could say that putting so
much power in the hands of a small venture-capital company
in California makes many of us feel uncomfortable."
But Rambus doesn't stand to remain small for long. A
back-of-the-envelope calculation suggests that if the DRAM
industry rebounds to $50 billion by 2001, and 60 percent of
DRAMs go to Intel-based systems, and 65 percent of those
DRAMs are Rambus types for which Rambus collects 1.5
percent, then Rambus earns about $300 million. In one year.
With no factories to build. And that's not counting royalties
from licensees such as Intel, surviving chip-set makers, 3-D
graphics-card vendors and non-PC vendors such as Nintendo
and SGI.
As Farmwald cheerfully acknowledges, "If you run the
numbers, Rambus stands to make a lot of money. Some
don't like that, but Rambus has done some enormously
important pioneering work."
Why wasn't that work done in Japan? Could Japan's research
labs have hit upon Farmwald's and Horowitz's radical ideas?
Not likely.
Farmwald says the Rambus concept sprang "from looking at
the memory bandwidth problem from a systems viewpoint. It
could not have come from a DRAM researcher." Horowitz
believes that Japan lacks the VC community and corporate
management that might support such far-out ideas. Others
note that the Japanese companies are notorious for poor
communications among their divisions, with computer
designers rarely discussing new ideas with the semiconductor
engineers, and vice versa.
Akira Minamikawa, senior semiconductor analyst at IDC
Japan, says that the rising importance of Intel and Rambus
has in part prompted Japanese companies to accelerate their
support for embedded DRAM technology and the hybrid
processes needed to fabricate memory and logic within the
same chip. Minamikawa credits Intel and Rambus for
prompting Japanese companies to support what had been an
academic research project at Kyushu University on the
parallel-processing RAM, in which multiple small RISC
processor and memory cores reside on the same chip.
PPRAM could one day be used in PCs, he speculates. "The
DRAM companies don't want to be reduced to being an
OEM manufacturer for Rambus," Minamikawa says.
Rambus's Tate argues that it is the DRAM makers
themselves who will benefit from the Rambus architecture.
By making desktops more attractive, everybody gains, he
says. "If DRAMs don't go up in performance, then PCs
won't sell. And the very first guys to benefit from Rambus
are the Japanese DRAM companies, mainly NEC and
Toshiba, which profit by being the leaders," Tate says. "By
being an early adopter, NEC has made a ton of money and
big strategic inroads at Nintendo, and they will keep
benefiting. The opposite is true for the companies that aren't
doing anything with Rambus technology."
Some fear that Rambus might "up the ante" in the future by
increasing its royalty and license fees when the industry is
"locked in" to using RDRAMs, Tate admits. But, he says, the
licensees "signed a semiconductor license, and nothing can
change the terms of the licensing. It is for current and future
patents, so even if Rambus acquires new patents, it can't
change our business relationship with that license partner."
Another source of concern is the gradual devolvement of
semiconductor content into an Intel microprocessor and
memory. Farmwald has ideas, which he stresses are separate
from those of Intel and Rambus. "Eventually, there will be
integrated MPU with memory, but that won't happen for 10
years or more on the desktop-though 10 years is an awful
long time in this business."
Farmwald expects some things to change in the early years of
the 21st century. Chip sets, which now principally control
graphics and memory, will move onto the microprocessor, he
suggests, particularly for low-end systems. Having a chip set
sit "in between" processor and memory slows performance
and adds cost. Thus, Farmwald says he wouldn't be surprised
if Intel some day develops an MPU with a Rambus interface.
Though Intel's official MPU road map ends well before 1999,
the expectation is that the first Rambus systems will use an
Intel chip set, a Pentium Pro (P6) generation processor and
RDRAMs. Not long after that, the P7 generation will begin
shipping. Meanwhile, the memory technology Intel hopes to
bring to the desktop is here now-in a $199 Nintendo game
machine you can buy at a toy store. Convergence, indeed.
-David Lammers is Japan bureau chief at Electronic
Engineering Times. He can be reached at lambo@twics.com.
Copyright r 1997 CMP Media Inc.>> |
