When Intel first launched its 64-bit Itanium microprocessor in May 2001, it did so with an uncharacteristic absence of fuss. Instead of the usual blanket marketing of consumer-targeted TV commercials, business magazine ads, and customer parties, the company quietly broke the news in a terse seven line press release. It stressed the new chip – the fulfilment of a decade-long push towards 64-bit computing – had only been unveiled to enable systems vendors to optimise their software for the new high-performance platform.
However, when vendors started to run tests on Itanium, it was clear why Intel had chosen such a low-profile debut. Not only was the processor slower than most of its principal competitors in 64-bit processing – Sun Microsystems' Sparc, Hewlett-Packard's (HP) PA-Risc and IBM's Power – it also ran at a slower speed than some of the company's own 32-bit Pentiums.
Unperturbed, Intel assured the industry that the follow-on Itanium would be very different, offering performance of between one-and-a-half and two times over the initial chip. In July 2002, Intel finally delivered.
For corporate organisations, Itanium 2 is a watershed – but not one that is painless. The sheer raw performance will enable them to start using Intel-based servers for high-end applications and databases that have historically been restricted to Unix servers. Moreover, the lower cost of the Itanium boxes, essentially for the same processing power, will likely force the Unix system vendors to lower their prices to match Itanium server levels.
But that assumes Itanium 2 has left behind all of its predecessors' shortcomings. For a start, there are some key architectural differences between Itanium and Itanium 2.
First, to reduce latency and improve performance, Intel has moved the 3 megabyte (MB) level-three cache from the motherboard and integrated it onto the Itanium die. The cache stores frequently used data. Moving it from a separate chip on the motherboard enables it to deliver data twice as fast and also reduces potential bottlenecks, says Intel.
Second, Intel's engineers have reduced the chip's pipeline stages and increased the number of execution units. This should help improve data throughput and enable it to handle more instructions per clock cycle when running computational intensive applications.
Third, Intel has tripled the system bus bandwidth from 2.1 gigabits per second (Gbps) to 6.4 Gbps, enabling a faster flow of data between the microprocessor and other components on the motherboard. Finally, the clock speed has been cranked up, but only from 800 megahertz (MHz) to 1 gigahertz (GHz).
At the same time – and perhaps equally significant – is the parallel development by Intel of a number of chipsets that ought to make it easier for hardware vendors to more quickly and cheaply put together their systems. The chipset governs the flow of data around the motherboard and traditionally, like the microprocessor, this would be designed by the systems vendor. The development of chipsets by Intel instead will bring a new PC-like level of commoditisation to the server market.
The first of these chipsets is optimised for 2-way and 4-way servers. But before the end of the year, Intel will provide support for 8-way and higher levels.
The upshot is that these off-the-shelf chipsets will enable systems vendors to put together larger and larger systems with more commodity parts, enabling them to offer increasingly powerful servers at more attractive prices.
Itanium's price/performance will present a challenge to the high end of the server market where Unix systems dominate – no more so than to Sun Microsystems. Sun is the only remaining major vendor to have not committed to migrating its server architecture to Itanium 2. HP is moving to Itanium from PA-RISC, and IBM is going to offer Itanium systems alongside its Power-based systems.
If Intel's claims that, architecturally and in terms of performance, Itanium 2 is already pulling ahead of Sun's UltraSparc III, then Sun needs to act quickly and aggressively if its RISC chip is not to run out of headroom.
Richard Dracott, group marketing director of the enterprise platforms group at Intel, claims that Itanium offers much greater system bus bandwidth, many more on-board registers and execution units, and can handle six instructions per processor clock cycle to UltraSparc III's four.
In real-world applications, Intel claims that this already gives Itanium 2 a significant advantage over UltraSparc III. In transaction processing applications, for example, Intel claims that Itanium 2 is 50% faster than the Sun chip. But architectural comparisons are complicated by the fact that, for example, Sun handles parallelism within its Solaris Unix operating system, rather than as a part of the core chip architecture.
In a bid to stay ahead of Intel, Sun has boosted its UltraSparc development group to 1,400 people, making it arguably the largest semiconductor research and development team in the world dedicated to a single chip design. "The Intel guys are working across Pentium, Celeron, Pentium III, Itanium, Xscale and so on," says Sun marketing manager Matthew Keep. "We just do UltraSparc."
The move also reflects Sun's desire to avoid another slip in the UltraSparc product roll out plans of the kind that saw the current UltraSparc III arrive one-and-a-half years late. If it can do so, the plan is for UltraSparc to match Itanium gigahertz-for-gigahertz.
UltraSparc IV will debut in the first half of 2003 and its speed will be boosted from 1.2GHz at launch to about 2GHz by the end of 2004, just prior to the launch of UltraSparc V in early 2005, which will run at up to 3GHz.
Into enemy territory
Yet at the same time, Sun has made some tentative moves towards the Intel market. In addition to its Intel-based Cobalt email and web server appliances, Sun will also release a range of low-end 'Lintel' servers – featuring Sun's own distribution of Linux on Intel. These are due in August 2002.
This is part of a broader strategy from Sun to embrace the open source operating system (alongside its own Solaris product) and, by implication, Intel. This is a critical step, says Philip Dawson, an analyst and program director of infrastructure services at the Meta Group. "Linux on anything other than Intel is really a non-starter."
The thawing of Sun's attitudes towards Linux is also evident in its porting of Linux application programming interfaces to Solaris and vice versa. This ought to make the migration or porting of applications in either direction much easier.
In contrast, following its acquisition of Compaq, HP has the unenviable task of assimilating seven different operating systems that it has historically supported and migrating them to Windows, Linux or HP-UX Unix.
But in the process, it has already infuriated many longstanding users of its proprietary HP3000 mid-range servers, by deciding to discontinue sales of the system's MPE operating system from October 2003. HP will, therefore, need to handle the migration of its community of users to Itanium-based servers with the greatest of care.
Although PA-RISC/HP-UX applications can be run on Itanium without recompilation or any other form of modification, they will need to be optimised if users are to take advantage of the extra performance that Itanium will offer – particularly when the PA-RISC roadmap draws to a close with the PA-8900 in 2003.
"It's probably a good idea for developers and software vendors to do a native port or re-compilation on to Itanium because RISC has no understanding of some of the architectural constructs within Itanium. Therefore, the code that was developed for PA-RISC cannot take advantage of it," says Chris Franklin, enterprise server marketing manager at HP.
What this indicates is that as systems vendors move to Itanium, some organisations will face a forced migration of applications and operating systems before they are convinced of any price/performance benefits.