"This is a demonstration that Intel is continuing to execute with precision, which is good" and very much a turnaround from the first-generation "Merced" version of Itanium, said Nathan Brookwood, an Insight64 analyst. "The Merced program was plagued with a variety of gremlins, but the McKinley and Madison programs have executed pretty smoothly."
Intel's ability to crank out new, faster versions of Itanium each year is a key part of its argument that it can use its manufacturing expertise to take on IBM, Sun Microsystems and Advanced Micro Devices, each of which have their own server processors. Intel must meet a higher standard for Itanium to catch on because adopting the new chip requires companies to completely rewrite their software.
"That's a pretty furious pace in the server business--a new part every year. It's much faster than what the other vendors are doing," said Kevin Krewell, senior editor at the Microprocessor Report.
Improving technology is critical for Intel, but it's not the only requirement for the ultimate success of Itanium. It will take years to develop all the necessary software for the Itanium line, which uses completely different instructions than Intel's Pentium family; the first major server version of Windows for Itanium is scheduled to arrive in April 2003. Some analysts expect Intel server sales tocompeting designs through at least 2007, while others expect Itanium chips will only start shipping in with Madison.
The Itanium processor line, co-developed with Hewlett-Packard, has a 64-bit design that can accommodate the vast amounts of memory needed for large databases. Other features protect against data transfer errors and make the chip well suited to large servers with numerous processors.
Pentium and Xeon processors, 32-bit designs, have a much smaller 4GB memory limit. AMD, whose chips are compatible with the Pentium line, is taking a different design course than Intel, working on a chip called Opteron that has 64-bit extensions but can still run all the software of the 32-bit family.
Intel said the size of each Madison will be 374 square millimeters, will consume a maximum of 130 watts, and will have 410 million transistors, the basic circuitry elements out of which chips are assembled. Until now, Intel hadonly that the processor would have "about half a billion" transistors.
Intel confirmed Tuesday that the chip measurements disclosed in the conference agenda are what's expected for the final version of Madison, expected in the summer of 2003.
"Those numbers are certainly what we expect and what we are targeting," Intel spokesman Seth Walker said. With the clock speed increase and the doubling of cache size from 3MB on the current Itanium 2 to 6MB on Madison, Intel expects overall server performance to increase about 50 percent compared with Itanium 2, with overall system costs staying about the same.
Big bucks for small chips
The 374 square millimeter "die" size is a bit smaller than the size of Itanium 2, but still is large and therefore pricey. The larger the chip size, the fewer chips can be carved from each of silicon.
"It's big," said Dean McCarron, a Mercury Research analyst. "But because it's a server product, that kind of die size is acceptable. It'll sell for thousands of dollars. With that kind of price tag, you can afford to have a really large die."
Intel declined to comment on pricing, but Insight64's Brookwood expects the cost to stay about the same as the $4,226 price of the current Itanium 2 and the $4,227 first-generation Itanium. "What they've done with the Xeon and Itanium line is that each new generation of chips is using exactly the same price," he said.
Server chips--larger and with more features--don't run as fast as desktop models. The Pentium 4 currently tops out at 3GHz, for example. But server chips have different design goals.
"Where you're focusing on a more complex architecture...you may not run it as fast, but you're doing more work per each cycle," McCarron said. Itaniums can execute about six instructions per clock cycle, compared with one or two for Pentium processors, he said.
The Microprocessor Report's Krewell, who expected Madison to run between 1.4GHz and 1.6Ghz, believes Intel may have been able to get Madison to run faster but would have run into problems with overheating that would have undermined one of the key selling points of the chip: easy upgrades.
Madison and its successor "Montecito" will both plug directly into servers that accommodate the current Itanium 2 processors. That makes it easier for server designers--of which there are--to build machines that won't be obsolete in a year. A key part of that promise is that the newer chips won't consume more power than the original Itanium 2.
"The goal is to show that within a fixed infrastructure of 130 watts, Intel can take it from McKinley to Madison," Krewell said.
Intel has said Madison will sport a variation of the Itanium 2 name.
The first Itanium 2 was built with a manufacturing process that used circuit features as small as 180 nanometers, or billionths of a meter. Madison uses the 130-nanometer process of the current Pentium 4, which permits more transistors for a given amount of silicon. Montecito, due in 2004, will use a 90-nanometer process still under development.
With the cache size and clock speed increases, Brookwood expects Madison to beat current Itanium 2 chips by about 50 percent on the much-watched SPECint2000 speed test.
AMD's Opteron, also due in mid-2003, will have a SPECint2000 score of about 1,202, AMDin October. The current 1GHz Itanium 2 posts a score of about 810, and a 2.8GHz Xeon tops the chart at 1,032.
With Madison and faster Xeons on the way, Intel should be able to meet AMD in the speed race, Brookwood said. "All three architectures will have very similar SPECint numbers," he said. "The two suppliers will be neck and neck."