If you can live with single precision floating point, a cheap GPU will crunch numbers about an order of magnitude faster than a CPU. That assumes the computation can be parallelized. If the computation cannot be parallelized then single threaded performance matters most. Single threaded performance differences between similar CPU's tend to differ more modestly and small performance improvements tend to come with much higher costs.
Depending on data size, more RAM and faster disks and faster networks may make more of a difference. But again, GPU's and parallelization are going to reduce computational time more than CPU choice when parallelization is an option.
Good luck.
