I finally done a library for Java with support of 256 bit integer arigthmetic, and other operations. Actually I did also 128 bit one.  These have support for both signed and unsigned numbers.
So some results

add 101 - 150 ms
sub simillar
div  between 10 s - 19 s
mul 1300 ms

All tested for 1000000 itereations.  Thus for example add is aprox up to 250 cycles.

I know I should post it in anoncements, but I would like to know if I didn't missed any algorithms that could speed things up.

For multiplication, I used the standard algorithm

q[0] = a[0] * b[0]

q[1] = a[0] * b[1] + a[1] * b[0]

and so on

For example 10 * 22
q[0] = 0 * 2
q[1] = 0 *2 + 1 * 2
...

For division, I used the algorithm with a subtraction and a bitshift.

I'm probably missing the point, but doesn't BigInteger cover this stuff already?

Why didn't I think of that?

That was kind of my next question. How do those performance figures compare to doing the same thing with BigInteger?

Edit:
It might also be intersting to compare against other math libs like apfloat (first one I encountered while googling on BigInteger performance).