This is an excellent piece of analysis. One question occurs to me: what will
be the overall context switch overhead costs by the increased FPU state save
and retrieval?

It seems to me you've gone most of the way in instrumenting some of the
costs. Do you think, given tne stats you currently have, that if you could
have usec differences between a context switch *with* an FP save/restore
vs. one without?

-matt