ScienceMark is a benchmark written by Tim Wilkens, which measures the CPU time necessary to perform advanced scientific calculations, and produces a score based on those times. It consists of three individual benchmarks; a molecular dynamics simulation involving liquid Argon atoms at 140 degrees Kelvin, one which finds all quantum solutions to the Schrodinger equation for all 61 electrons in the element Promethium, and a calculation of the total energy in a water molecule using Quantum Monte Carlo (QMC).
The first two (actually first and third) tests in the ScienceMark suite operate almost completely independently of memory bandwidth. This is a good example of why some programs scale better than others. In the molecular dynamics and QMC tests, the data set being manipulated fits entirely within the 256KB cache available to the processor, and as a result, scales linearly with processor speed.
The Primordia test, however, illustrates a different picture. It is clearly more dependent on memory bandwidth, as the two test configurations with the least amount of bandwidth turn in the slowest numbers, and indeed, both begin to scale poorly around the 1.8 to 2 GHz mark. The 500/333 platform, courtesy of plenty of available bandwidth, is able to scale well up to, and beyond, the 2.0 GHz mark.