Next week, Intel Corp. will roll out the long-anticipated and much-delayed Itanium architecture, its first commercial iteration of the IA-64 family, with an 800-MHz processor code-named Merced. And though large system OEMs are ramping two-, four- and 16-processor computers for imminent introduction, many of them are opting to wait a little longer for the superior McKinley chip, the next Itanium generation, to launch more complete product lines.

"This will be a three- to five-year ramp," said Mark Hudson, worldwide marketing manager for the HP9000 product line at Hewlett-Packard Co., which helped develop the Itanium architecture and is in the midst of readying two-, four- and 16-way systems.

Merced is shaping up as a kind of development platform, as OEMs await the more capable McKinley — due next year — to design their systems around. "The first version of IA-64 will be targeted at several key areas, pilot programs and early developers," Hudson said. "When we get to the McKinley time frame, that will be a major boost. McKinley will round out the product line."

"McKinley is slated to have first silicon by the end of the year and systems by this time next year, so it makes more sense to wait," said Kevin Krewell, a senior analyst with MicroDesign Resources (Sunnyvale, Calif.). "HP and others will support IA-64 but wait for the more stable McKinley platform."

HP will continue to support and invest in its PA-RISC solutions for the next five years, Hudson said, while migrating its customers over to the Itanium platform. One of HP's earliest contributions to the architecture lay in making that migration process as smooth as possible.

"We've done a lot of work to make it a viable transition," Hudson said. "You can literally move your application running on PA-RISC today over to the Itanium processor offering without any change." HP unveiled a new generation of PA-RISC solutions last year, most likely due to the continual delays in the Itanium rollout.

For its part, Dell Computer Corp. this week offered reporters a glimpse at the four-way PowerEdge 7150, its first server based on IA-64 and the 460GX chip set from Intel. The machines will be rolled out over the summer, with similar offerings due from IBM Corp., Compaq Computer Corp. and Silicon Graphics Inc. this summer and fall.

Mike Fister, vice president of Intel's enterprise platform group, said Intel has sent out about 7,000 Itanium-based systems to users for software testing. Some 30 OEMs plan to introduce Itanium-based systems, he said, including "mainframe manufacturers in Japan" that Fister did not name. When McKinley ships in 2002, and the Madison and Deerfield processors ship the following year, usage will broaden gradually to Internet service providers and other midrange server applications, Fister said.

Industry analysts generally see the McKinley as a chip worth waiting for. "This Merced is the first iteration, and Intel is still working out some of the details," said Krewell of MicroDesign Resources. In developing the architecture, Intel deferred many improvements to the second generation, he said. "Itanium is still a new architecture, with new software running on it, and it will take a while to get accepted. As they delved more into the architecture, Intel found things to improve that were too late to add to Merced. So they pushed those improvements off to McKinley, which will be a much better product."

For stability's sake, Intel will manufacture both chips on an 0.18-micron process, Krewell said, making the migration from Merced to McKinley smoother, on a proven manufacturing line. This will shorten the qualification time for McKinley, keeping the product line at a fixed, constant design. "But the next step, 0.13 micron, is where the architecture will really start to take off," he said.

Floating-point clout

Analysts praised Merced's floating-point performance numbers in early benchmarks, but termed the integer benchmarks ho-hum. "When it comes to floating-point, numerically intensive applications, this might be the fastest chip that anybody's ever made — excluding Cray, but that's a different league," said Nathan Brookwood, president of market research firm Insight64 (Saratoga, Calif.), referring to the supercomputer company.

"Merced's performance is not as exciting when it comes to general-purpose kind of benchmarks involving integer calculations," Brookwood said. "But in addition to the floating-point numbers, Merced brings to the party the ability to handle much larger programs and databases than any of the 32-bit chips."

For its part, "McKinley will come in at much higher clock rates, to at least around 1.2 GHz," said Linley Gwennap, president of market research firm The Linley Group (Mountain View, Calif.), and it will also solve some of Merced's chief bottlenecks. "Then Intel will take McKinley into 0.13 micron as soon as it can."

In Gwennap's view, "On basic, integer kind of code, standard stuff, Merced doesn't measure up very well. Itanium has a very small Level 1 cache, and that could be the problem on some of these benchmarks. Intel will fix that. And Itanium has a relatively slow bus to the main memory, so Intel will fix that too." Indeed, he added, "the Pentium 4 has higher bandwidth to the main memory."

Analysts argued that the Pentium 4 ramp has stolen some of the thunder from the Itanium performance numbers. "For applications that can live within the constraints of 32-bit architectures, the Pentium 4 P4 is still the chip of choice," said Brookwood, though Merced "would have been clearly outstanding" had it appeared on its original due date, in 1998.

"In the meantime, the P4 clock rates are far in excess of what anybody anticipated. For general-purpose applications, the P4 at 1.7 GHz remains the fastest guy on the block," he said.

Merced, however "was really targeted" for applications like transaction processing that demand more memory than the 4 Gbytes at which 32-bit architectures cap out, Brookwood said. Gwennap agreed, and said, "To a certain extent, Intel is Intel's biggest competitor."

The ever-shifting timetable for Itanium — which Intel had once planned for 1998, then pushed out till 1999, then 2000 — has effectively rendered the Merced chip an evaluation platform for future systems to many OEMs.

"If it was out in 1999, that would've given the industry two years to assimilate Merced-based systems," Brookwood said. "The delays have shortened the period the Merced chip is in the market. A lot of folks have concluded that they will use these initial systems to pilot IA-64-type applications, and deploy them only in small volumes. With McKinley around the corner, it makes more sense to hold off."

Krewell likened the situation to the Pentium architecture's adoption rates. "The first Pentium-architecture chip was expensive and had a few problems, and the next generation, Pentium II, was much more successful," he said. "It's the same pattern here."

"Merced won't develop that whole IA-64 platform, so the next generation will develop it more fully," concurred Mark Briscoe, a product-marketing manager for Tektronix Inc. The test-and-measurement equipment provider works closely with Intel to support the chip maker's next-generation platforms. "There are always interim steps."

Cooling challenge

According to Dell, one of the biggest challenges for Itanium-based systems going forward will be how to dissipate heat from the processor module. An 800-MHz Itanium consumes about 130 watts — almost double the power consumption of a Xeon processor, also running at 800 MHz — partly because Itanium has double the cache of a Xeon.

When McKinley is introduced, processor speeds will increase, making the power consumption issue more problematic. The launch of chip sets that support eight-way and up to 32-way configurations, and the arrival of rack-mounted servers with denser Itanium boards, will make the challenge even more critical.

Bob Van Steenberg, vice president of Dell's high-end systems, said that the four-way system on display last week used conventional fans running at higher rotation speeds than usual. However, Dell engineers are considering blowers that would send cooling air directly over the processor modules, he said. But then, high-powered blowers require ducts to direct the airflow, complicating the issue.

"We are talking to several emerging startups that have some innovative solutions," Van Steenberg said. "Our research people are doing some mock-ups and testing the heat-dissipation capabilities. And Dell Ventures has made some investments in companies that have technology in this area." He declined to be more specific.

In keeping with Intel's revolutionary approach, OEMs will have to do a lot of redesigning and reengineering to move their product lines over to the Itanium architecture. Those not willing to walk that road may look elsewhere — to Sun's Ultrasparc or Advanced Micro Devices' Sledgehammer lines.

"The end user will have to go through some transition, and it's not going to be trivial," said Gwennap of The Linley Group. "But once there, they'll have a wealth of choices in terms of OEMs and OS support, which isn't something you'll get out of Sparc."