Intel's new Itanium chip drags its feet

Most Itaniums are expected to ship at the lower speed, Mike Fister, general manager of Intel's enterprise platforms group, said during a news conference at the Intel Developer Forum. Intel will start selling the chip in the fourth quarter of 2000, said spokeswoman Christine Chartier.

The chip's speed is a lesser factor than architectural improvements such as the 64-bit design that allows it to hold vast databases within memory, said Insight 64 analyst Nathan Brookwood. But the slower speed indicates difficulties with the manufacturing process for the large new chip.

"Clearly there's a problem there. You always love for the distribution to favor the high end," Brookwood said. However, he added, "I don't think the megahertz rating on the initial Itanium is going to be very important in the broad scheme of things."

Intel servers are growing up but still have a long way to go before they can be used to attack the entire product line of Sun Microsystems, which has become Intel's biggest rival. The delayed Itanium chip has been the flagship of Intel's effort to unseat Sun's dominant 64-bit UltraSparc-based Unix servers.

Intel emphasizes several uses for its Itanium chip and successors in the IA-64 line. Among those uses are encrypting and decrypting information, analyzing large quantities of business data and housing large databases, Fister said. Currently there are about 6,000 prototype Itanium computers being tested, with chip speeds generally running at 500, 666 and 733 MHz, he said.

While IA-64 is the point of Intel's attack on Sun, Intel is expected have more immediate success with extensions to its current 32-bit line.

In servers, the 32-bit line is topped by Xeon chips, which feature larger on-board high-speed "cache" memory than ordinary Pentiums and the ability to be grouped into servers with as many as four processors.

Just as a Pentium 4 is coming out this fall, a high-end Xeon equivalent code-named Foster also is due, Intel has said in the past. Today, though, the company talked about another new model, called Gallatin.

Gallatin is built on the 0.13-micron manufacturing process, a smaller feature size that enables Intel to build chips that are faster and less expensive. It's the high-end parallel to Northwood, the successor to the 0.18-micron Pentium 4.

Like Xeon, Gallatin will have a range of cache sizes, Fister said. The first Gallatin chips will come with smaller caches, with larger caches coming later. But where Xeon chips top out at 2MB caches, Gallatin will be larger, Fister indicated.

Brookwood said 4MB would be the likely cache size. "That's a lot of transistors," he said. Current large-cache Xeons have 140 million transistors, the tiny electronic components that make up chips, Fister said.

IBM's upcoming Power4 chip, which crams two CPUs in a single chip along with a large cache, uses 170 million transistors.

But CPUs are only half the game when it comes to building servers, Fister said. Where desktop computer performance is strongly influenced by the CPU horsepower, servers require fast connections to memory and input-output systems for network cards and storage. The chips that handle these tasks are called chipsets.

"It's like a race car. If you drop a 1,000-horsepower engine into a Yugo, you'll burn the tires off it," Fister said.

For Foster workstations, which are essentially gussied-up desktops, Intel will offer the 870 chipset, Fister said. But for Foster servers--expected to be a much larger market--Intel will rely on ServerWorks and other third-party manufacturers, he said.

However, Foster servers won't debut until the second quarter of 2001--a full quarter after the workstations, Fister said. The first Foster servers will be dual processor-capable machines, with four-way, eight-way and higher arriving in the third quarter of 2001.

Concurrent with the arrival of Gallatin in "early 2002" will be Intel's reentry into the server-chipset market with a new system called Plumas, Fister said. This chipset will enable two-processor servers and is timed to take advantage of mainstream adoption of the InfiniBand high-speed system to connect CPUs with storage, networks and other servers, Fister said.