Hammer strikes a blow at Itanium
AMD's new 64bit processor may present a serious challenge to Intel.
Will AMD's new 64bit processor present a serious challenge to Intel, which is attempting to attract manufacturers of high-end servers to its competing Itanium chip?
AMD has been reasonably successful at making processors for desktop PCs, especially for home users who are more sensitive to price differences.
To date, however, it has not managed to make significant inroads into the professional server market, even at the lower end of the spectrum.
But the firm has recently been talking about its forthcoming 64bit processor family, codenamed Hammer, which could upset the progress of Intel's 64bit IA-64 architecture. Indeed, AMD is making some claims for Hammer that bring into question Intel's strategy in the high-end server market.
One of the problems, both for vendors and IT managers, is that the definitions between low-end, mid-range and high-end systems are becoming blurred.
For example, it is increasingly likely that the mid-range or even high-end requirements of companies will be supported by a cluster of low-end systems. Dell, for example, already markets a supercomputer constructed in just this fashion. And IBM's supercomputer offerings are now all clustered eServers.
Smooth transition
Against this background, AMD's new Hammer processors could offer a smoother transition from the current 32bit Wintel server to the world of 64bit systems. The Hammer design remains close to the old x86 architecture of Intel's renowned Pentium, Celeron and Xeon devices. Indeed, the official designation of the processor, x86-64, points clearly to its lineage.
This means - or at least AMD claims - that Hammer will be able to run existing 32bit Wintel server applications a good deal faster than a 32bit processor.
The same cannot be said of Intel's Itaniums. These processors will be able to run existing 32bit applications, but will not offer any genuine performance advantage. In fact, Intel warns that Itanium is likely to run such applications slower.
It is true that most important enterprise applications have already been ported to the 64bit IA-64 platform, so there is an argument that suggests the route through to a 64bit environment is well covered. This is certainly the case for those users who know exactly what applications they wish to port to IA-64, but some remain uncertain and may find that Hammer is a more attractive option.
Applications
An Itanium-based system requires not only investment in hardware and operating systems, but in new applications. It will also require the inevitable commissioning processes before it is allowed to go live. It is possible to predict, therefore, that the take-up of Itanium systems will, of necessity, be rather sluggish, and will largely be restricted to pilot systems for some time to come.
Intel's own product roadmap suggests that it recognises this. There is clear delineation between Intel's IA-64 processors and the continuing 32bit Pentium Xeons, for example. New implementations of the Xeon are planned until 2003, so the company does not intend to drop its 32bit interests.
While this may provide a product set that is straightforward for Intel to manage, it is not necessarily such a neat solution for its customers.
Organisations currently using Wintel servers will want to move from 32bit to 64bit systems at different times for different applications.
IT managers, already facing higher expenditure owing to Microsoft's forthcoming licensing changes, may be uncomfortable if they have to pick either 32bit or 64bit systems - many might prefer an option that allows them to use both.
AMD's Hammer processors might offer a more attractive transition path. It should provide a performance improvement when running existing 32bit applications, and the same hardware can be used for the transition to a 64bit environment.
Advantages
Hammer may have additional advantages. One is its memory architecture - cache-coherent Non-Uniform Memory Access - basically a Numa-based processor on a chip. Numa can already offer faster memory access in systems design, especially in large multiprocessor environments such as IBM's Sequent systems, and the approach is now coming in at chip level.
Early examples, however, have suffered from variable latencies in fetching data from memory, but AMD has said that this problem is fixed.
So Hammer is now tightly coupled to local memory and provides a fast access speed regardless of the number of processors involved. A four-way system has a claimed latency of 140ns, for example, while an eight-way system has a claimed latency of 160ns. This should help server manufacturers with multiprocessor designs.
Another factor that could make Hammer an attractive option for enterprise users is the way in which its multiprocessing facility makes it relatively easy to build fast systems that scale using either symmetric multiprocessing or clustering technologies. This could be a real advantage for IT managers trying to decide when and how to step up from 32bit to 64bit systems.