Concerns about the performance of Intel's 64bit Itanium chips have resurfaced following publication by Top500.org of new supercomputer benchmark results that rank the Pentium 4 chip faster than the Itanium 2.

Top500.org is the leading publisher of supercomputer performance measurements, and its tables of test results are frequently quoted by computer makers to help promote their products.

However, it seems unlikely that Intel will rely on Top500.org performance data to support its Itanium range for the foreseeable future. This is because Intel's consumer-oriented 32bit Pentium 4 outpaces the firm's high-end 64bit Itanium 2 chip by seven per cent in Top500.org's latest main Linpack benchmark league table.

However, IT Week Labs' analysis of the results highlights the importance of using software that is specially compiled for the processor architecture that is being tested.

The Pentium 4 leads the Linpack benchmark results because the benchmark's rules forbid the use of optimisations for a particular processor architecture. Tested under those conditions, the 2.53GHz Pentium 4 is seven per cent faster than the top-of-the-range 1GHz Itanium 2 chip.

However, a better indicator of chip performance comes from the Top500.org's related "best effort" Linpack test results, because this test allows for processor-specific optimisations. The results from this test show the two chips in a radically different light - indicating that the Itanium 2 is 20 per cent faster than the Pentium 4.

This discrepancy between the two sets of results highlights the dilemma facing many IT managers considering the Itanium architecture. While the Itanium's ability to run specially tuned software is impressive, its efficiency drops dramatically when running software that is optimised for a different architecture.

A straw poll by IT Week Labs indicates that currently, most open source business software is not optimised for any architecture. And most commercial software packages are optimised for Intel's 32bit Pentium 3 chip.

The Top500.org data also provides some useful benchmarks from other high-end computing systems. A 16-way configuration of NEC's SX-5 chip leads the best effort tests with a score that is some eight per cent higher than its nearest rival, which is an eight-way system that uses the latest version of the NEC chip, the SX-6.

The best score by a high-volume server architecture was for a 16-processor IBM pSeries 690, which is ranked some 40 per cent slower than the best.

The Linpack benchmark is widely used, but it is important to remember that it is based on systems solving complex mathematical problems. This is a very different workload compared with those used for transactional database benchmarks, such as those published by the Transaction Processing Council.