binary-trees is rough to beat C on. It involves allocating and traversing some pretty gigantic data structures, and the C code is optimized by using region-style allocation to avoid all of the individual calls to free (or garbage collections in a GC'd language). From the CPU usage, it also looks like they also haven't parallelized the Fortan version yet, though even at ideal speedup it'll still lose.

I've been doing some tuning work on our parallel dialect of ML, and many of the C programs are fairly well-optimized, though mandelbrot still has some room for improvement.

From a simple look, the Fortran code seems to have an additional optimization that allows it to avoid a lot of checking for overflows in the innermost loop.