The current generation of ARM processors have terrible floating-point performance. IIRC, the Cortex-A8 and Cortex-A9 are nine cycles for a double. The latest Cortex-A15 is finally single-cycle double, but is not yet common in the market. (EDIT: See Addendum below)
Certainly, in the embedded world where I belong, a 400MHz PowerPC - theoretically two doubles per cycle - (Freescale MPC5121e) outperforms a 800MHz Cortex-A8 (Freescale i.mx535) by about 50% total time (including IO) for a numerically intensive application. The PowerPC outperforms the ARM by a factor of 2.5x for the actual numeric part! And the ARM has the benefit of 4+ years of kernel and GCC improvements!
A modest Intel Atom N450 1.6GHz netbook outperforms the 400MHz PowerPC by a factor of 4x, both in total time and the numerically expensive core.
So no, I cannot recommend ARM cores for numerical applications, at least not for several generations.
I was a little wrong on the Cortex-A9. Whilst it is indeed a nine-cycle double, unlike the Cortex-A8, they can be pipelined and hence get results close-ish to 1DMFLOPS/MHz. The Cortex-A8 is barely more than 0.1 DMFLOPS/MHz (though single is a different story).
ATLAS benchmarks for the Cortex-A9 can be found here. Benchmarks for the Cortex-A8 can be found here.