Problems with numerical accuracy?

[Vegard Aaker]

 Hi,

I am writing a optimization algorithm that solves a quite large linear
system of floating point numbers. The algorithm works well on a PC,
and should be quite robust to floating point errors. But when I am
using a microcontroller the algorithm crash, most likely due to
numerical issues.

I am using the Ethernut 5 board and Nut/Os that use newlib as the
floating point library. Now I am wondering what kind of accuray I can
expect from newlib compared to for example a PC (running windows)?
Should I expect much more accuracy if the microcontroller contains a
FPU?

Kind regards
Vegard Aaker