dmf and Thread - General Interest Article on Intel-HP IA64 partnership
eb-mag.com
The birth of a new processor
Inside the INTEL-HP MARRIAGE: cooperation, squabbles and mysteries
By Russ Britt, illustration by Gary Baseman
It was an auspicious beginning. Perhaps suspicious is a better word.
Flashback to 1993. Both Hewlett-Packard Co. and Intel Corp. are looking for a way into the 64-bit chip market. HP wants to make its servers run faster and better compete with the likes of Sun Microsystems Inc. and IBM Corp. Intel, on the other hand, is looking for a little respect outside the safe confines of the desktop-processor market. They both have ideas, but they need help carrying them out. Each gets wind that the other is working on something.
The first step was to make friends with the enemy?a sensitive mission. HP decides to make the first move, and it goes slow at first. Jerry Huck, HP's chief architect for the project eventually named Merced, says the company tried to look for insiders who weren't too well-entrenched. "What we'd try to do is find allies inside Intel that had [worked] somewhere else."
Then in late 1993, HP and Intel decide to exchange notes, says John Crawford, Huck's counterpart at Intel. It wasn't exactly the most relaxed atmosphere, he recalls. Into the first three months of 1994, the companies did a lot of "peeking under the covers" to see what each organization had, Crawford said. They picked a neutral site that was halfway between Crawford's group in Santa Clara and Huck's in Cupertino. It was considered hallowed ground that the two firms agreed to keep neutral, even though it was a building owned by HP. Half a dozen engineers from each company met at the facility every few days.
"We couldn't bring anything into the room. Well, we could bring in slides and stuff, but we couldn't take out any notes," Crawford said. He says the two companies kept all the discussed material in a locked file cabinet at the site. "If we had decided not to cooperate, I guess we would have blown up the file cabinet," he said.
They did cooperate and formed their partnership later that year. But the chip, now known as Itanium, remains a source of contention between the two companies, and a cause for concern among analysts. Market observers worry over the chip's viability?even as it prepares to debut later this year. Some say it's underpowered compared to soon-to-be-released competitive chips.
"I wouldn't call it TENSION, but it required more effort than what we're used to. Any RESISTANCE was more along the lines of whether a feature was GOOD OR BAD." ?Jerry Huck, chief Itanium architect at HP
Also, analysts say the underlying issue is whether Intel and HP did what few technology companies have done before in situations like this: Made their marriage work. How have they nutured the relationship and how strong is the bond? And are Itanium partners and customers worried about troubles in the marriage and what that might mean for them?
Potential threats
Itanium faces a number of potential threats that could seriously impair its market acceptance. Consider that Itanium already is officially a year behind schedule?some say it's longer than that, as HP and Intel originally planned to have it out in late 1998.
There's more, though. HP is hanging on to its PA-RISC line of processors longer than some thought it would, and has balked completely on using the first version of Itanium for one line of its servers.
Many companies involved in the project are lowering the lifeboats by making contingency plans for other business, just in case Itanium sinks like a brick. One partner in the program points out that the massive cartridge housing the chip and its cooling system?it looks like a brick, in fact?could be too big for some companies to use.
Concerns are mounting so much that the joke making the rounds at a recent trade show was that Intel picked the wrong name to take Merced from concept to creation. Instead of "Itanium"?intended to remind people of the resilient metal titanium?they say the chip maker should have called it "Itanic."
All it took with the Titanic was "one iceberg to prove them all wrong," said one chip industry observer. "I don't know if there is an iceberg out there. But that has to be (Intel Chief Executive) Craig Barrett's nightmare."
"If we decided NOT TO COOPERATE, I guess we would have blown up the file cabinet." ?John Crawford, chief Itanium architect at Intel
The bottom line is that Itanium probably won't deliver the goods everyone thought it would bring when it still was called Merced, a chip that would instantly dominate the server market and claim it for its own, says Linley Gwennap, server analyst for Cahners' MicroDesign Resources. He says Itanium is unlikely to deliver a performance advantage more than 30% greater than Compaq Computer Corp.'s Alpha, today's speed champ. While that seems hefty, Gwennap says it may not be enough.
"Is that enough, in and of itself, to put everybody out of business? I don't think so," Gwennap comments.
But Nathan Brookwood, an analyst at Insight 64 in Saratoga, CA, says just beating out Alpha or even the Sparc processors made by Sun may be enough to put Itanium over the top. "If Intel can take on those guys and come out with a better product, that will be very compelling," Brookwood says.
Marriages
The history of the industry is littered with past failed joint efforts to build a microprocessor, most notably the IBM Corp.-Motorola Inc.-Apple Computer Inc. plans for the Power PC microprocessor. That partnership never really made a dent in Intel's stranglehold on the desktop chip market.
Gwennap says partnerships often lead to bottlenecks in the bureaucracy, particularly when a large company is involved. He notes that with some of IBM's partnerships the computing giant often has had two managers signing off on every decision, no matter how trivial.
HP and Intel tried to avoid most of that while developing Itanium. The two companies were 50-50 partners as the blueprints were drawn up, but they held roles designed to complement each other, Gwennap says. HP basically came up with much of the original design, while Intel turned the dream into reality by virtue of its manufacturing muscle. In short, HP was the brains and Intel the brawn.
THE MERCED/ITANIUM PROGRAM
products, estimated release dates and features
Version Due out Features
Merced/Itanium Q3 2000 800 MHz, processed in 0.18 micron
McKinley Late 2001 Processed in 0.18 micron
Deerfield Mid 2002 Cost-reduced McKinley
Madison Early 2002 Processed in 0.13 micron, should have 0.10 micron capability by 2004. Speed should be 3 to 4.5 times that of original Itanium.
N/A Late 2004 New core of 64-bit chip
SOURCE: CAHNERS MICRODESIGN RESOURCES
Since then, HP has faded into the background. Intel has assumed more control over the processor family and will sell the chips. But HP hasn't faded away completely, and it's led to some friction. Gwennap says he's sensed this, and sources in the industry back up that claim.
"Both sides say the same things about each other," says one chip industry source who's close to the program. "They both take the credit, and they both think the other gets too much attention." Ultimately, Itanium is Intel's chip to market and sell. The powerful chip maker already is beginning to flex its might. The source says Intel is beginning to make harsh demands on some of its partners on the project and threatening to back away if its wishes aren't granted.
Gwennap says Intel has been talking tougher to HP, too. "Intel is taking a lot of the credit for it," Gwennap says. Engineers and others connected with the project for HP are frustrated and annoyed, Gwennap says. Executives say they don't care about the credit, but there does seem to be increased tension between the two, and there's already been some fallout from that tension.
HP backs off
Take HP's decision to bypass the first version of Itanium for its mid-range "N" class servers. The move has left some in technology circles to tout the first version of Itanium as nothing more than a test model. HP says it will wait until the follow-on, code-named McKinley, shows up sometime in late 2001 to put it in those servers. Intel officials say they are puzzled by the decision. HP is putting Itanium in its lower-range servers to start.
Despite the confusion, analysts say it might make sense for HP to wait since the first version of Itanium will be on the market barely a year before McKinley is ready. Yet some of HP's other moves seem puzzling to Intel. HP has extended its PA-RISC line, not long after the company claimed Itanium was its future in servers, according to analysts and other sources. Gwennap says the last PA-RISC model was supposed to be the 8700, but the line was extended to the 8800 and 8900 when, in mid-1998, Intel pushed back Itanium's release date from 1999 to mid-2000.
Gwennap says HP had promised at one point not to continue building PA-RISC past a certain point. One chip industry source says Intel is furious that HP chose to step back from this plan. "This has created tension between the shops because HP is looking for an incremental advantage while [the two companies are] also trying to align," he says.
It's also unclear what kind of licensing agreement HP and Intel have forged, if any. The companies won't say whether HP gets any royalties from each chip sale. Analysts doubt Intel would part with any of its highly prized operating margins, which have exceeded the 30% level on occasion. What is more likely is that HP gets a first peek at all future versions, so it can get a jump on competitors in hardware design, analysts say.
But many involved in the project say HP probably is making a prudent move. Customers will need more time to transition to Itanium and its descendants.
Hal Feeney, an analyst with Pathfinder Research in San Jose, CA, says anyone thinking about shifting to Itanium might follow HP's lead. "I think for any company, as they're making a change in (computer) architecture, they have to have some sort of back-up plan," Feeney notes.
COMPETITORS WITHIN AND WITHOUT
Company Chip Due Out Features
Intel Willamette Second half 2000 Seventh generation CPU core 1-GHz in 0.18 micron
Intel Foster Late 2000 1.2 GHz
Compaq Alpha 21264 Now shipping 1.1-GHz model due out in first half of 2000
Compaq Alpha 21364 Late 2000 New system interface; 1.6- GHz version due in early 2002
Compaq Alpha 21464 2002 New multithreaded core
Sun UltraSparc-3 First half 2001 600-MHz at 0.18 micron
Sun UltraSparc-4 First half 2001 1 GHz, 0.15 micron, copper
Sun UltraSparc-5 First half 2002 New core, 0.13 micron, copper
IBM IBMPower3 Now shipping Out in 200-MHz. 500 Mhz due in 2000.
IBM IBMPower4 2001 2 PowerPC cores on chip
SOURCE: CAHNERS MICRODESIGN RESOURCES
But Itanium, whether it be the first model or follow-on versions, should gain acceptance over time, Feeney says. He points out that the architecture should have a lifespan of 10 to 15 years. That gives Intel plenty of time to build its market share.
Merced's genesis
The Itanium saga began in the late 1980s, when HP began thinking it would have to find a way to move from 32-bit processors for its servers to 64 bits. RISC architecture used to build HP's older lines of PA-RISC chips wouldn't cut it in a 64-bit setting, says Scott Emo, an HP marketing manager. "It's like trying to build a brand new house on an old foundation," he said. So the company needed to dream up a new chip blueprint.
One of the first concepts HP looked at was very long instruction word, also known as "Wide Word." It was sort of an introductory course in the bachelor's program for parallelism, or having a processor handle a multitude of instructions at once. Emo says there were two problems with Wide Word. For one, it wasn't scalable.
Second, the processors could handle eight instructions at a time?no more, no less. That means all software had to send out instructions in batches that were divisible by eight. If programs didn't do that, they had to be expanded in order to bring in the right number, creating a lot of wasted space, Emo says.
So in the early 1990s, HP started looking at ways to go a step beyond Wide Word. It began gravitating toward a concept known as explicitly parallel instruction computing, or EPIC, sort of a cross between RISC and Wide Word. It was around this time that the company realized it needed help to develop this technology and bring it to the market. It needed a powerful ally.
At the same time, Intel was looking into 64-bit architectures. It sought the means to break into higher-end servers but didn't have experience designing a high-end chip. So the two companies decided to get together.
Culture clash
From the beginning, it was clear the two companies needed to cross a cultural bridge. Intel had its way of doing things and HP?well, it had the HP Way.
"I wouldn't call it tension, but it required more effort than what we're used to," says HP chip architect Huck. "Any resistance was more along the lines of whether a feature was good or bad."
Intel architect Crawford remembers the cultural differences being a little more pronounced when it came to making decisions. "The difference with HP is that they were much more consensus-oriented. They discussed it until everybody agreed," he says. "Where at Intel, if we'd take the consensus approach, great. If it went too long, a decision-maker would say, we're going this way... HP would say, 'Let's work it out.'
"I think it led to some confusion. We tried to accommodate it as long as we could," Crawford says. "We tried to modify the decision-making process in a way each of us was comfortable with." Crawford adds: "For the most part, we were able to get a good working relationship from the start."
That culture clash didn't throw the Merced project off track, though. Ron Curry, Intel's Itanium marketing director, says it was more Intel's determination to avoid other collaborators' past mistakes. The company borrowed a page of sorts from the EPIC parallel design and decided to build many parts of the chip simultaneously.
However, by not building the chip in a sequential fashion, many components had to be redesigned, Curry says.
"It didn't work. As a result, we went back to our best-known methods... and applied those methods. We [then] said it would take six to nine months longer than we thought," Curry says.
ITANIUM SYSTEM MAKERS, PARTNERS
As of November 1999
AST ICL
Compaq NCR
Data General NEC
Dell Olivetti
Fujitsu Samsung
Gateway Siemens-Nixdorf
Groupe Bull Toshiba
Hitachi Unisys
HP VA Research
IBM
SOURCE: CAHNERS MICRODESIGN RESOURCES
For some on the project, though, turning Merced into Itanium has taken an inordinate amount of time. "I would have never guessed it would take this long, back in '94 when we started," says HP's Huck. He says the time it takes to get to market on a new chip design should take roughly four years, two-thirds the time it will take Itanium.
Huck maintains, however, that the chip is precisely what he envisioned. "It very much is what I wanted in the sense of an architecture."
Intel's Curry adds the results will be well worth the extra time taken. "The analogy is you can take a 1976 Chevy and put a new muffler on it, but it's still a '76 Chevy. Or, you can buy a new 1999 titanium-alloy Ferrari."
Rivals line up strategies
While Itanium may be worth the wait, the delays in getting Itanium to market are giving would-be rivals more time to line up countermeasures.
Sun already is campaigning against Itanium. The Palo Alto, CA-based workstation and server maker is the only major hardware company that won't be using any type of Itanium architecture in its machines. The company circulated a pamphlet at last year's Microprocessor Forum titled "A Few Things You're Not Being Told About Merced." It posed several questions such as "How long will customers have to wait before Merced ships in fully integrated and debugged systems?" and "Can 32-bit applications all be ported to IA-64 without rewriting any code?"
"We believe that customers who are contemplating a transition to the Merced architecture have to transition to a lot of things," said Steve Fritz, Sparc marketing campaign manager. Sun claims Itanium won't be as scalable as its own systems. And Sun officials point out they've had 64-bit architecture on Sparc for some time.
Insight 64's Brookwood says Itanium may require an updating of some systems, but it is more compatible with a broader array of software and hardware systems than processors made by Sun. "If you go down the Sparc path, you are married to Sun," Brookwood said.
Compaq is proposing new architectures for Alpha, due by 2002, that the company hopes will answer the Itanium threat. Two designs in particular are being featured?simultaneous multi-threading and time-division multiplexing. The first involves spreading a microprocessor's power across several functions, the second is like having two processors divide tasks, Brookwood says.
The strategy, however, makes Alpha somewhat of a moving target, he says. In some cases, the processors have to add a scheduling function so they don't duplicate tasks. Itanium's EPIC design avoids that, Brookwood says.
One potential threat
There might be a dark horse in the 64-bit processing race, though, he adds. Advanced Micro Devices Inc. is developing a chip code-named Sledge Hammer. Unlike Itanium, Sledge Hammer is an x86-based design. Fresh off a victory of sorts with its new Athlon processor, Sunnyvale, CA-based AMD might score another coup in the workstation and desktop market if it can get a chip out fast enough, Brookwood says. By sticking with the x86 design, AMD enables users to keep certain programs that might not migrate to Itanium. More importantly, it allows system developers to remain with a processor design that's familiar to them.
"I think it's a very clever application on AMD's part," Brookwood said. He points out AMD, Intel and Compaq-Alpha all could be on equal footing in 2001. That's when all three companies will have new server-class chips on the market in their relatively early stages. "AMD has some significant advantages because its 64-bit platform will be (close to) its 32-bit platform," he says.
One by-product of Sledge Hammer is that Windows-based operating systems are a good bet to be the software of choice for Sledge Hammer. That could pit Microsoft Corp. against its longtime ally Intel. It should be said, though, that Windows NT also is expected to be used on Itanium.
And there's always the same concerns with any processor coming out of AMD. The company's manufacturing shortcomings have caused trouble when it's tried to deliver the goods in the past. Steve Lapinski, AMD director of product marketing, says the company needs to make sure it produces what it promises. The chip is due out around 2001.
Given the fact there have been delays, tension between partners, skepticism in the technology community and enemy ships approaching, getting a hold in the high-end server processor market won't be automatic for Itanium and Intel.
"[Itanium is] probably going to be [in] one or two systems [that each enlisted company offers.]," says analyst Gwennap. "It's not going to go across all their product lines." |
