In the messages I said that I increased number a bit and the result became even worse. That is consequence of bad asymptotic, so further increasing leads to further increasing of difference. But the goo implementation must be a lot faster than java's as it uses Karatsuba algorithm that is faster than the current KMath's naive one but is a lot slower than FFT based ones (like in haskell).

I've merged benchmark with some not quite correct changes. The problem is that KM BigInt does not support pow operation, so I am not sure that any measurements with it are correct. By the way, on GraalVM, basic add and multiply seems to be much faster on KM than on JVM

Or not... at least it is similar

Large number multiplication seems to consistent with your tests

In order to make comparison for power valid, one need to implement power in a non-recursive way.

I was quite surprised by having naive implementation for power in the corresponding file. I'll fix it.

I fixed the base in dev, but it is still too slow

Also, I think that power that are outside ULong are practically useless but lead to overhead. Negative, negative powers are now forbidden.

BigInt cannot implement PowerOperations as root cannot be defined and non-natural (or 0) powers are forbidden for it.

(Added

pow

to my fork)

Added Karatsuba: https://github.com/mipt-npm/kmath/pull/328