Ok, so I tried compiling Numpy in scratchbox. I was happy to find that the sourcecode ships with rudimental blas and lapack implementations in c. Using these numpy compiles without problem. As do matplotlib. So it is actually easy to have a lightweight numpy/matplotlib on the IT. That more or less eases my need, since that means I can make applications which relies on linear algebra and plotting but not on extreme performance.

Of course it would also be great to have the real lapack for it's efficiency. I actually tried to compile clapack and cblas. It worked and the test program passed most tests (it is apparently normal for a lapack installation to fail some of the tests. Simply because they are compiled on a different compiler)

However, it is almost 10 Mbyte library.