Data link layer ?

One of the difficult things about trying to quantify the effect of adding a numeric co-processor is that not only do different applications use numeric co-processors to different extents, but a single applications use numeric co-processors to different extents, but a single application will vary in its use of a co-precessor depending on the job it is tackling. For example, it is generally held that spreadsheet programs are heavy co-processor users but it isn’t difficult to find examples of spreadsheets that do virtually no floating point arithmetic! For example, a spreadsheet that has been constructed to act as a database is very unlikely to perform any arithmetic worth installing a co-processor for.

To find out the likely impact of a numeric co-processor, estimate how much time your program spends in floating point calculation. Even in situations where the application appears to do nothing but arithmetic the estimation can be much more difficult than you might expect. For example. Suppose you have a spreadsheet consisting of numeric data and formulae and it takes 30 minutes to recalculate. A simple minded approach would suggest that if you add a mumeric co-processor that does floating point calculations five times faster than software the calculation would only take six minutes. If you try this out you will usually be disappointed. The reason is that the 30 minutes to do the calculation isn’t occupied solely by floating point calculations-data has to found and results returned to the correct locations. In all probability no more than 75% of the 30 minutes is spent in performing floating point calculations and this is the only portion of the calculation time that can be reduced by the addition of a co-processor. In this case the calculation time would go down to only 12 minutes instead of the promised six.