Analysis As expected, Intel effectively sidelined its decade-long, multi-billion-dollar VLIW processor by announcing 64-bit extensions to its IA-32 processor family today.
Itanium represented one of Intel's biggest gambles: the company had never successfully moved its customers from the x86 instruction set onto an entirely new instruction set. Although the underlying architecture of Intel's processors has changed radically, the instruction set has essentially remained the same, taking evolutionary steps forward while retaining backward compatibility.
But the VLIW (very large instruction word) architecture also posed a radical challenge for software developers, as it put the onus of optimizing the code onto the compiler writer. If the compiler did a poor job of second-guessing how the code should best run, then software would run slowly, no matter what was added to the hardware - usually in the form of expensive large on-die caches - to improv3e performance.
Still, when the Itanic first set sail it seemed like none of these problems would be insurmountable.
The project began at Intel in December 1991 and was announced as a partnership with Hewlett Packard - which had also been investigating VLIW designs - in 1994. Intel assigned its best designers and a bottomless trough of marketing dollars to the technology.
By 1999, while the world was still waiting for real silicon, Intel had persuaded the most important operating system vendors to port to IA-64. Microsoft, SGI, Compaq, HP, Novell, Sun Microsystems, IBM (with SCO and Sequent), and the Linux Trillian project all announced or showed native IA-64 code that year. And analysts were almost unanimously agreed that Intel's volume production of IA-64 would destroy the economic justification of the development of rivals such as Power and SPARC.
"In all my years of following this industry," wrote John C Dvorak, "I have never seen such sheeplike behavior."
Five years later, the level of software support for the Itanium tells its own story. Microsoft half-heartedly markets a version of Windows for IA-64, while Monterey, Modesto and Solaris are mothballed, and only HP has a strategic interest in migrating its HP-UX and Tru64 customers onto the processor, as HP and Compaq discontinued development of their own Alpha and PA-RISC processors.
Two generations on, IA-64 has failed to provide the promised volume economics. See, for example, Dell 14, IBM 0 - quarterly Itanic sales revealed. Of 160 million PCs shipped last year, the IA-64 was in just 3,000, although with a Biblical loaves and fishes trick in December, Intel's Paul Otellini somehow turned that 3,000 into a feast for 100,000.
So where now, for Itanium?
Intel has made expensive gambles before. IA-64 is in some sense, simply history repeating itself. In 1981 Intel bet the farm on its iAPX-432 'Micromainframe' processor.
Work on this very advanced chip began in 1975, and the team was full of good ideas. The chip incorporated object-orientated design principles - it was designed to be programmed in the Ada language - and had multiprocessor support built in. Great attention was paid to fault tolerance.
However, it was much slower than the x86-compatible 80286, and proved to be large and expensive, with three (and eventually, four) chips on the package. Post-mortems cited "code-optimization issues" blamed on "compiler weakness". All very familiar issues for today's Itanium coders.
The i432 was eventually sidelined in 1986, the year Intel's 80386 processor appeared. The 80386 was backward-compatible but added more conventional 'mainframe' features such as the ability to run virtual x86 machines. The x86 instruction set didn't take such another quantum leap forward until AMD, realizing that Intel was struggling to move the mass market to Itanium, saw an opportunity with x86-64.
The i432 project cost Intel more than a billion dollars. But valuable lessons were learned, and fed through to Intel's successful embedded RISC processors, for example. Bob Colwell, chief architecture honcho for the chip that saved Intel in the mid-1990s, the P6 (Pentium Pro), described the i432 as "a wonderful research project masquerading as a bad product".
(When the P6 appeared Intel was under immense pressure from RISC vendors such as IBM, MIPS and to a lesser extent Sun. The P5 Pentium had failed to land a knock-out blow, but the P6 saw Intel vault to the top of the performance ratings, outpacing even Alpha.)
Five into two will go
Ironically, key engineers of the P6 were rewarded by being reassigned to work the IA-64 salt mines - just one of the opportunity costs that Intel will have to calculate when the Itanium project is said and done. On the plus side, the historical determinism (and kindergarten economics) that informed what Dvorak called "sheeplike behavior" - the mantra about volume economics - has had some upside for Intel.
The mere threat convinced MIPS and HP to put their RISC processors into maintenance mode, Compaq CEO Mike Capellas to axe Alpha, and Sun to look for partners such as Fujitsu to continue SPARC development. So while Itanic may be one of the computer industry's most expensive flops: it has reduced the RISC competition from five to two.
So is Itanium dead? Not at all, says Intel, and the company can cite work on the multi-threaded multi-cored Tukwila (formerly Tanglewood) processor which is expected in two years' time, and which we exclusively revealed here.
Miracles have punctuated Itanium's history before (see Miracle cures Berkeley man of Itanic wickedness and Analyst sees St. Fister in Itanium wafer, not to mention the Feeding of the 100,000), and may happen again. However to pay back its decade-long investment, IA-64 now needs a turnaround of miraculous proportions. ®
Intel's Barrett hints at 64-bit compatibility glitches
Intel unveils 64-bit capable Xeon
Intel's Otellini promises 'Year of Itanium'
Intel toasts Itanium's success by giving servers away
Analyst sees St. Fister in Itanium wafer
Miracle cures Berkeley man of Itanic wickednessAMD claims to win battle of slowest selling chips
Intel and IDC at odds over Itanium's future
Multicore Itanic: Call me Tukwila
First dual-core Itanic to sport 24MB of cache
Intel's Tanglewood pumped full of DEC Alpha goodness
Intel Tanglewood's first voyage