Usage of the AI server as "general purpose" server

Hello!
I have a question that is very interesting as for me. Let explain that we have some software. And this software can use MANY threads in its work. For example, calculation of the salary for several thousands of people. And lets imagine that we also have some AI-specialized hardware. As I understand, this hardware is a combination of the lots of simple calculation cores. And they can make calculations in the parallel mode.

Here is the question. Is it possible and reliable to use such hardware (hardware Neural Processing Unit) to solve some simple mathematical problems that can be paralleled to the lots of threads?

I think that if we will try to do something like that, we will need to manually calculate the weights of the links between “neurons” to force them to do exactly that operations that we need. I am not sure that it is possible. And I am not sure that it is good idea. My suggestion is that in this scenario GPU calculations will preferred – they probably easy to program and debug.

Does anybody have some other opinion?

Thank you!