[necessity]

>BTW, do you have numbers for setups with hundreds of GPUs?

I saw a talk on it a while ago, I can only remember they were using CUDAlign and Smith-Waterman (the basic idea is the same). Doing some googling too this seems to be a reasonably recent work with GPUs and CUDAlign (DOI 10.1109/CCGrid.2014.18).

>I'm also left wondering about results using stochastic solutions.

Another talk, I think they were running Smith-Waterman too. The speculative part was during the traversal of the matrix to get the edit distance. It's not the most time consuming part of the algorithm. I got in late for the talk and I didn't get to hear what they did about filling the matrix in the first place, but I imagine they might have done something similar. I'm not very familiar with Smith-Waterman so I can't go into details.