Great if you've got a suitable problem
As usual, these things are great if you have a problem which is a good fit to what it can do. In the science world, that would be either something which spends all of it's time inside BLAS or FFT, or at least has a compact, highly-vectorizable AND highly-concurrent kernel at it's heart.
If, on the other hand, you have a problem with a flat execution profile (so that you'd need to rework all the logic and rewrite 1e5 LOC to take advantage of a GPU), and, god forbid, long dependence chains everywhere (so that you gain no advantage no matter how much you rearrange things), the whole thing becomes a very expensive equivalent of a vintage Celeron.
I know Cray was ridiculed for his "If you were plowing a field, which would you rather use: Two strong oxen or 1024 chickens?", but I think a balanced diet is essential for long-term health - and lately, all we've been having is chicken.
Biting the hand that feeds IT
