Pick Your Title: (1) A Case In Point, (2) Why Merced Is Vaporware, Just Like NT 5.0, (3) Hewlett-Packard Earnings Announcement - Revenues
Up Only 4% (and I don't wonder why) (4) Another Gigantic Opportunity For Sun Microsystems or (5) Why HP Will Lead Tech Stocks Down Tomorrow, Not Up
Ken Wilson aka High-Tech East
Intel and HP: Strange bedfellows; Different Agendas Strain 64-bit Chip Development Plans; Merced Won't Be Expected 'RISC killer'
By Lisa DiCarlo, PC Week Online
November 16, 1998 9:00 AM ET
Two years ago, engineers at Hewlett-Packard Co.'s development labs
came to a painful conclusion: The 64-bit Merced chip HP was developing with Intel Corp. would not be powerful enough to surpass HP's existing PA-RISC architecture. So, HP set out--on its own--to design a new
processor based on the same underpinnings as Merced but with a different system I/O and bus implementation as well as other capabilities. When Intel learned of the project, it convinced HP to extend its IA-64 partnership, which originally covered co-development work only on Merced, to include the new chip. That's how Merced, once considered a potentially crushing blow to the RISC community, has evolved into a mere steppingstone toward the goals Intel and HP originally intended for it when they announced their historic partnership in 1994. Merced's focus will be on backward compatibility with Intel's X86 architecture. The follow-on chip--which Intel has
code-named McKinley--is the one that will go head-to-head with high-end 64-bit RISC architectures.
That's bad news for HP as well as the myriad other systems makers lining up behind IA-64. Merced is not due to ship until mid-2000, and McKinley isn't due until late 2001, putting Intel about three years away from delivering a 64-bit architecture that can comfortably surpass the performance of RISC systems. HP executives won't comment on their involvement in McKinley's development.
The importance of Intel and HP successfully delivering a powerful 64-bit platform to the rest of the computer industry cannot be
understated. Everyone has a stake.
Almost every systems company will build IA-64 servers, and all major Unix operating system developers plan ports to IA-64. Some companies, such as Silicon Graphics Inc., will eventually abandon their own RISC
architectures in favor of IA-64 running Windows 2000 or some flavor of Unix. And Wintel stalwarts such as Dell Computer Corp. need a solid 64-bit platform to compete for future enterprise customers.
"All of us expected the first [IA-64] chip to be hard, and it has been," said Jim Davis, general manager of the IA-64 program office at HP, in Palo Alto, Calif. "You learn everything on the first one, so the second and third ones get significantly better."
Problems with Merced should surprise no one, considering the unlikely
partnership between Intel and HP. The two industry giants are working
together toward a common goal but approach the partnership with vastly
different agendas: Intel's lifeblood is selling millions of compatible processors, while HP needs an architecture to help customers migrate from its aging PA-RISC platform. And in this partnership, Intel holds most of the cards. Terms of the agreement stipulate that Intel has control and final approval over the chip's design and implementation, according to several sources. As a result, Merced will have the required backward compatibility with X86 applications. That's a critical feature for the millions of Intel customers but is not likely to convert RISC system buyers looking for higher performance.
"What [Intel was] doing [with Merced] wasn't what we needed," said a
source close to HP. As a consequence, "Merced won't look so great
against PA-RISC."
When Merced debuts in mid-2000, it will have a clock speed of about
800MHz. By comparison, the 21264 Alpha from Compaq Computer Corp.
is expected to reach 1GHz by that time.
Merced's other technical advantages, which are largely dependent on
somewhat obscure instruction-level parallelism in software, are not
expected to outweigh Alpha's clock speed advantage, according to Linley Gwennap, vice president of MicroDesign Resources Inc., in Sunnyvale, Calif.
As it stands now, Merced won't offer a compelling technical advantage
over IBM's Power3 or Sun Microsystems Inc.'s UltraSPARC processors
either--assuming those chips continue to evolve as scheduled, Gwennap
said.
Intel officials in Santa Clara, Calif., dispute claims that Merced will fall short of performance expectations. Merced will be "very competitive" with RISC chips, said Ron Curry, Intel's director of marketing. "Will McKinley blow your socks off? Yes, but Merced is pretty good, too." But that's assuming that those products are not late to market, as they have been in the past.
Reluctant partners Merced performance is not the only issue on which HP and Intel disagree. The two were reluctant partners when, driven together by outside market forces as well as their own product development and financial shortcomings, they decided to work together in June 1994. Finding a successor to PA-RISC was HP's intention in 1989 when it began developing the EPIC (Explicitly Parallel Instruction Computing)architecture, then known at HP as WideWord Architecture. At the time, HP believed that the RISC architecture, including its own PA platform, would eventually hit a performance wall.
"Anybody who thinks RISC will go on forever doesn't know what they're
talking about," said Joel Birnbaum, senior vice president of R&D at HP
Laboratories. But when HP completed that 64-bit architecture in 1993, executives realized that the company could not shoulder the skyrocketing development and fabrication costs of a new architecture. Furthermore, the potential for a voluminous Unix software base was quickly slipping away.
For the first time, HP considered relying on another company to meet its high-end processing demands. "It was a difficult decision," said Lew Platt, HP's president, chairman and CEO. "It was clear that we needed a partner that would be able to develop and deliver these chips in volumes higher than HP could." Intel, the chip company with the deepest pockets, was in a precarious position of its own in 1994. Its 2-year-old 64-bit effort, known as P7, was having problems. Intel later abandoned P7 for EPIC, although some properties of P7 were incorporated into Merced.
A bigger problem was the looming threat of PowerPC, a RISC architecture announced by IBM, Motorola Inc. and Apple Computer Inc. in October 1991. At its conception, PowerPC posed a major potential
challenge to Intel, since it was being designed to run Windows NT, Unix and Macintosh operating systems.
"The [PowerPC] project had everyone's attention," said a source who
was closely involved in the HP negotiations for Intel. "[Intel and HP] had a common foe in terms of a new RISC architecture."
So, HP approached Intel in early 1994 and proposed that the two work
together.
Intel was elated over the prospect of having HP-UX, an enterprise-level, scalable operating system, available for the first time on its architecture.
"That was a major attraction," said the Intel source involved in the talks. "We contemplated other choices, and this one went down very easily." Those other choices included a partnership with IBM on PowerPC, according to sources. Intel and IBM discussed some kind of collaboration over PowerPC, although the two sides never came close to an agreement.
HP and Intel signed their deal in June 1994, but not before HP made
major concessions.
First, HP had to agree that Intel would control all of Merced's design
decisions, even those involving the EPIC architecture that HP had already developed. For HP engineers who had worked on the architecture for more than five years, this was the most bitter pill to swallow.
Also, HP could not publicly discuss the work it was doing on Merced and other IA-64 chips, nor could it sell its own IA-64 processors, even though it jointly owns the instruction set architecture with Intel. "It was a big blow to our egos to turn over control [of the EPIC] architecture to Intel," said Birnbaum, an 18-year HP veteran who was a driving force behind the Intel relationship.
The muzzle Intel put on HP regarding Merced made HP vulnerable to
companies such as Sun that were proactively poking holes in HP's
strategy.
"Our competitors were out there giving road maps to 2050, and we
couldn't say anything about what we were doing," Birnbaum said. "We
finally told Intel, 'This isn't fair.'"
Intel has since relaxed that restriction, enabling HP to provide long-term system road maps to customers.
To make matters worse, HP engineers decided that Merced performance
would fall short of expectations. "We reached the reluctant conclusion in 1996 that we would have to design another chip" for HP-UX systems, the HP source said.
That's when HP set out independently to design a new chip. Enter Intel, again.
"This [McKinley] chip achieves a different objective than Merced, and
[Intel] became very interested in it," the HP source said.
A new deal HP was willing to strike a second deal for the same reasons it went to Intel for Merced. And once again, it made major concessions. Just as HP couldn't comment on Merced, it could not publicly discuss the new chip, even though it was being developed largely at the HP Labs facility in Fort Collins, Colo., with Intel engineers on site, according to sources.
"Intel wants to pretend [it's] designing the chips alone," MicroDesign's Gwennap said.
Intel provided the first technical briefings on McKinley early last month and disclosed further details of the chip later in the month at the Microprocessor Forum. When asked then about HP's involvement, Intel declined to discuss the company; it continues to downplay HP's
involvement with McKinley.
"It's an Intel product," said Albert Yu, senior vice president and general manager of Intel's microprocessor products group.
"HP brings a lot more than casual comment [to the development], but
we're working with many hardware and software companies," added Yu,
a 22-year Intel veteran and one of the architects of the HP partnership.
Analysts expect McKinley--which gets its code name from the highest
mountain in North America--to compete with higher-end RISC
architectures, while Merced--named after a California river--will target RISC workstations and entry-level and midrange RISC servers.
"[Merced's] primary goal is compatibility, and performance comes with
the next generation," said Martin Reynolds, an analyst at Dataquest Inc., in San Jose, Calif.
Some vendors said that they cannot predict Merced's performance, but
they're concerned that its delivery timetable will result in less-competitive products, as RISC vendors such as Compaq, Sun and IBM enhance their chips.
"You lose your performance advantage every month you're late," said
John Kelly, vice president of server development at IBM, in Armonk,
N.Y.
Merced systems were originally expected in 1997 or 1998; Intel has
already publicly announced a six-month delay for Merced, pushing it into mid-2000, and has missed internal targets for packaging issues, support circuitry, architecture and the like, sources said.
"No one will acknowledge [the internal delays] because, technically, it's not a slip until you publicly have to restate your schedule," said a source familiar with the development timetable.
The delays have frustrated HP, partly because some setbacks resulted
from problems in areas where HP's expertise exceeds Intel's--problems
that might have been fixed sooner.
Intel had its own issues early on with HP, which quietly talked up the
architecture's performance despite its pledge not to discuss the project and the fact that a viable product was years away.
"This was a contentious point," the Intel source said. "It wasn't time to talk about the capability because [the instruction set] wasn't even defined."
The two companies continue to walk a fine line in their relationship: HP is itching to get more credit for the development of Merced and McKinley, while Intel does not want to alienate its OEMs by creating a perception that it is favoring HP.
"This was a big concern about two years ago, but if [OEMs] weren't
reasonably convinced that we've created a level playing field, they
wouldn't all be participating," Intel's Curry said.
HP, for its part, does not want to continue in its chip-maker role beyond McKinley. Instead, it is turning its focus to compiler technology and sophisticated system design, through which it hopes to gain an edge over the throng of companies also working on IA-64 servers and workstations.
That's a role that seems to suit both sides well.
"We can focus our energy on developing chips, and [HP can focus] on
systems," Yu said. "We made investments in big wafer fabs, and HP
preferred not to."
The basic architecture upon which IA-64 is based is expected to last
about 20 years. HP and Intel's co-development contract has no hard
cutoff date, so it's possible the two could collaborate again--one center stage, the other in the shadows--on post-McKinley projects.
And despite HP executives' desire to get more credit for their
behind-the-scenes work, at least one ranking official understands the role Intel must play in promoting its next-generation platform.
"The future of IA-64 will depend on people seeing this as an Intel-led, Intel-driven standard," Davis said. "Working with them was our heart's desire, and we got lucky." |
