End of the road in sight for Windows on Itanium

It was perhaps one of the most drawn-out, painful launches in Intel's long history: the introduction last February of Tukwila, the latest generation of its Itanium 64-bit processor architecture. Not everyone in the Itanium Solutions Alliance hung on for the five-year ride, with Unisys having been its most prominent drop-out last year, citing competitor HP's dominance in the field. Microsoft held on for the entire stretch; but last week, the company announced it would not lend its support to whatever the generation after Tukwila might become.

In a blog post last Friday, Windows Server Senior Technical Product Manager Dan Reger said that his company will continue to support existing Itanium architecture, including Tukwila, for another eight years. But Windows Server 2008 R2 will be the last version of the operating system to support IA-64.

After a delay of officially two years, and unofficially longer, the dual- and quad-core 65 nm Itanium 9300 series was launched last February 8 to somewhat muted fanfare. The newest Itaniums ended up following, not leading, many of Intel's latest process innovations, with the company's 32 nm Xeon 5600 series, including a six-core model, premiering just last month. Itanium can no longer claim power savings as a selling factor, with the new Xeons also boasting 130W TDP at clock speeds superior to the new Itaniums'.

Five years ago, the drive toward industry-wide adoption of IA-64 architecture was started by the Itanium Solutions Alliance, steered somewhat by Intel (of course) as well as HP, the architecture's most prominent vendor. But Microsoft was also a very vocal charter member -- at the time, seen as more influential than charter member Red Hat. The reason was that, from 2001 up until 2005, the factor keeping enterprises from investing in Itanium was the lack of a clear migration path. Windows compatibility was perceived as the substance of a bridge.

Here's how I covered the Itanium Solutions Alliance in a January 2006 article for Tom's Hardware:

The Alliance's efforts go to the heart of what had been characterized as Itanium's chief deficiency: not its architecture, which after a rough start, actually has proven itself very capable. It is an altogether different platform, not because it's truly 64-bit, but because it would have its developers embrace a concept called Explicitly Parallel Instruction Set (EPIC). It is the industry's first CISC-based multithreading architecture, based on a simple concept to explain but a difficult one to implement: the idea that when processes fork and parallelism begins, it's because the code of the program tells it to. It's this concept which most starkly distinguishes Itanium from x86 (x64) architecture, which actually has no parallelism principles of its own. The ability of multithreaded x86 processors to fork processes into separate cores is based mainly on their ability to ascertain for themselves when such forking is permissible. One of Intel's most ambitious tests of this capability has been hyperthreading, which is a parallelism technique for single-core processors. But HT is an experiment that will probably come to an end as dual-core and multicore processors become mainstream. As they do, they will undoubtedly bring Intel's version of 64-bit x86 architecture (EM64T) as well as AMD's (AMD64) into the mainstream even in high-performance categories.

I was wrong about one thing: Hyperthreading does live on in Intel's current Core microarchitecture. Nevertheless, the fork in the road we saw in early 2006 has only grown wider today, and three major influences have changed the scenario for Itanium over that period:

• Linux is finding a comfortable home among high-end servers, because it's lightweight and has lower up-front costs (assuming businesses don't buy premium support licenses), and Intel is embracing Linux more and more.
• The hyperbolic growth in virtualization means that enterprises don't have to run Windows Server just to enable their clients to run Windows 7. Exchange, Office Communications Server, and SharePoint remain important, but any more, Windows Server exists to support them rather than the other way around.
• Heterogeneity in server design means businesses that require high-performance computing servers can now purchase Itanium in more limited quantity, leaving Windows Server to handle high-availability segments of the network.

In some ways, many of the factors standing in the way of the Itanium Solutions Alliance's goals five years ago, have actually been removed through the natural course of the industry's evolution. But that evolutionary course has not exactly favored Microsoft. So it's noteworthy that, in announcing its "transition" away from Itanium support last Friday, Microsoft's Reger (or at least, the blog post from Reger, whose opening had one tone and his closing another) took a parting shot at Itanium, aiming squarely at what the Alliance had touted as one of the architecture's key features: scalability.

"The natural evolution of the x86 64-bit (x64) architecture has led to the creation of processors and servers which deliver the scalability and reliability needed for today's 'mission-critical' workloads," Reger wrote. "Just this week, both Intel and AMD have released new high core-count processors, and servers with 8 or more x64 processors have now been announced by a full dozen server manufacturers. Such servers contain 64 to 96 processor cores, with more on the horizon. Windows Server 2008 R2 was designed to support the business-critical capabilities these processors and servers make available. It supports up to 256 logical processors (cores or hyper-threading units), so it's ready for the ever-increasing number of cores."

Or to put it another way, six is greater than four. Of course, Reger's comment (perhaps intentionally) ignores one important element: After Itanium finally included the QuickConnect memory bus (the equivalent of AMD's DirectConnect architecture, developed years earlier), and with Itanium's brand of parallelism being implicit (built into the code) rather than explicitly stated as with x64, three dual-core IA-64 processors should bear little performance difference from one six-core processor.

In a May 2005 interview, HP's representative to the Alliance, Stephen Howard, told me IA-64 architecture was noteworthy for having "very high-RAS features: reliability, availability, scalability. Those hooks are built into the chip architecture itself, and the operating systems can make use of those features, and the individual companies and software vendors can build on the most highly reliable system. So you get up into the mainframe replacement and high-end RISC replacement classes of servers, and anywhere that this kind of equipment is used is going to be an ideal spot for Itanium."

At present, Microsoft remains a member of the Alliance, and may conceivably continue to do so until the next edition of Windows Server, perhaps in two years' time. But one of the Alliance's goals, Howard told me five years ago, was to create a single point of contact for customers wanting to purchase Itanium hardware and software all in one stop. That isn't exactly what happened, but perhaps Intel's continued embrace of Linux, along with Red Hat's continued membership in the Alliance, could actually promote that goal in Microsoft's absence.