Pentium III defies the laws of physics

Amid the horror of Caminogate and Rambus, an increasing number of observers are singing the praises of Chipzilla's earlier chipsets, the desktop BX and the server GX. They may not be state of the art, but they have a major benefit: they work. In the case of the GX, they appear to work better than conventional wisdom would allow. Intel's Lancewood server mobo is a dual slot 1 device with all the usual server gubbins thrown in - LAN, graphics, etc. Having two CPUs is usually reckoned to give between 1.5 and 1.75 times the performance of a single processor. Why is it then that our reference PIII 500 machine, using a Sun River SR440BX, returns a Wintune 98 integer score of 1453.077 Mips and the Lancewood reports 2916.869 Mips rather than a best case scenario of two times the uniprocessor score? Both machines have 128MB of RAM, all three PIII 500s are the same stepping and have identical 6GB hard disks, so where do the extra Mips come from? And the mystery deepens. Original thinking was that the results were caused by Wintune 98 being confused by dual-processor configurations, but this is apparently not the case. A third machine has a Coppermine 700MHz, 256MB and a much faster, 20GB hard disk. It takes an average of between 10.5 and 11 hours to complete a SETI@Home work unit, with Windows 2K reporting 100 per cent CPU utilisation. The Lancewood takes between nine and ten hours to complete a SETI unit, Win2K indicating that only one processor is used to handle the task, using 95 per cent CPU time, the second processor merely ticking over at two to three per cent. So we have a 700Mhz machine with twice the RAM running flat out, being soundly thrashed by just one 500 chip in the dual box ambling along at 95 per cent of full speed. We always liked the Lancewood for its solid build quality and value, but now it appears to be wringing impossible performance levels from a couple of venerable 500MHz PIIIs, we like it even more. What's going on? Suggestions on a postcard please. ®

Related Story Leak! Intel's server board strategy to 2002