Itanium: Sun Microsystems' worst nightmare?

Although it doesn't blow the doors off the competition--as Intel had initially hoped (and hyped)--Itanium is a solid product, best viewed as the first in a long line of processors that will eventually dominate the high end of computing, pushing aside current competitors such as IBM's Power family and Sun's SPARC processors.

Ever since Itanium was first revealed, there has been much controversy about its likely performance. The chip's unique architecture, which Intel and HP call EPIC, is based on VLIW technology that is fundamentally different from the RISC and CISC technology used by all other high-end processors. This technology is designed to achieve higher performance but relies heavily on the compiler (which creates Itanium code) to anticipate and plan for all possible program outcomes. Critics wondered if the compiler could be this efficient.

With the Intel announcement comes the first official benchmarks on the long-awaited processor. These measurements, frankly, are a mixed bag.

Itanium excels on some benchmarks and does poorly on others.

Champ or chump?
At least some of these shortfalls, however, do not reflect fundamental problems. Itanium does best on scientific applications with large data sets and lots of complex floating-point math. Groups such as CERN (the European particle-physics laboratory) and NCSA (National Computational Science Alliance) love what Itanium does on the most difficult computational problems in the world. On some of these problems, it significantly outperforms even the Alpha processor, widely regarded as the world's fastest chip.

Itanium also does well on transaction processing and other high-end business applications. Based on preliminary TPC-C measurements, Itanium performs as well as the best available processors, including Alpha and IBM's Power. This performance is critical when running Oracle or similar database software.

The chip does not do as well on general-purpose business code, such as simple Web serving. On SPECint, a popular benchmark, Itanium is 30 percent slower than the fastest processors, including Intel's own Pentium 4 and HP's PA-8700. On this type of code, Itanium is hampered by its small caches, a shortcoming which Intel plans to fix in a future version, and by the inability of the compiler to always predict what these programs will do.

Intel has touted Itanium's compatibility with software that runs on its popular Xeon processors. In theory, Xeon customers can buy Itanium systems without needing to upgrade their software. In reality, customers will quickly discover that they have purchased a very slow machine. Intel won't even discuss Itanium's performance in "compatibility mode," but it is certainly slower than a Pentium 3 Xeon, much less Pentium 4.

Because of these issues, I do not expect Itanium to be broadly deployed. There will be some end-user deployment in the scientific segment, but this is a low-volume segment. In the commercial segment, most businesses will purchase only a few Itanium systems for software development and testing.

Broad deployment is likely to wait until the next-generation McKinley processor is available. McKinley, likely to be branded as Itanium II, is already being tested in Intel's labs and will appear in systems next year. Intel says this chip will deliver twice the performance of the original Itanium and have three times the memory speed. This big jump in performance comes from locating and uncorking several bottlenecks in the Itanium design that were not anticipated.

Further performance improvements will come when McKinley is moved from a 0.18-micron process to the 0.13-micron process that Intel will begin producing later this year. The 0.13-micron McKinley, code-named Madison, will be three times faster than Itanium and is expected to appear in systems in early 2003.

These rapid performance improvements will show the true value of Itanium's EPIC technology. McKinley and Madison should outperform all competitors on most benchmarks, even the troublesome SPECint. On scientific and transaction-processing code, the next-generation Itanium chips will be head and shoulders above the competition.

While these chips are cleaning up the high end, Intel also plans to address lower-cost servers and workstations. The initial Itanium already has a low-cost version at a price of $1,177, low enough for some workstations. The company is developing a part code-named Deerfield that should allow pricing well below $1,000, making it suitable for even low-cost workstations and servers. This chip should appear in systems by early 2003, about the same time as Madison.

Once Madison and Deerfield are in place, the Itanium line will be poised to entirely displace the current Xeon family. Thus, while Itanium will likely contribute only a few hundred million dollars to this year's top line, Itanium family revenues should exceed $6 billion by 2004. This will begin to help Intel recoup the estimated $1 billion-$2 billion that it has spent to develop Itanium.

Most server vendors have adopted a two-track strategy, offering customers a choice of their proprietary architecture or the new Itanium platform. Itanium systems will be available from more than 20 vendors this year, using popular operating systems such as Windows XP, Linux and AIX. With this approach, Itanium sales should offset any cannibalization of the proprietary RISC customer base. In the Itanium space, server vendors will differentiate on the basis of system performance, availability, support and similar factors.

The big loser in this transition is Sun, the only major server vendor that continues to use a single-track strategy, offering only its proprietary SPARC platforms and its proprietary Solaris operating system.

Even if Sun can match the performance of the Itanium family with its SPARC processors, it won't be able to match the system-level features of 20 Itanium OEMs. With Itanium, end users will have a choice of buying from many system vendors, all selling compatible systems. With SPARC, end users must buy from a single vendor.

I believe this business model will restrict Sun's business, causing a significant loss of share within two to three years to Itanium supporters such as HP, IBM and Compaq Computer. The company may ultimately counter by offering its own Itanium systems, but at that point, Sun will be decidedly late to the party.