One thing that could be done without a lot of effort is compile a bunch of those
example applications that come with the MS compiler using both it and the Borland
one. While not many of them are speed tests, one could at least get an idea of how
big the executables generated by each compiler are without worrying about the
VCL's overhead - some are WinAPI, while others use MFC, but that comes with both
systems anyway. Another thing to consider is DirectX: the SDKs come with the
source for some utilities that are meant to test the speed of graphics hardware,
so it might be interesting to see if compiling them with MS/Borland results in
different figures (MS supply libraries for BCB, so there shouldn't be too many
problems).

Another idea might be to write some examples that thrash the NAG C++ numerics
library, taking into account not only speed, but also the accuracy of the results
produced by both compilers.

If nothing else, fiddling with all this stuff might produce some optimisation
settings that help users of both compilers get the best possible results for their
applications.