node17 A SIMD Interpreter for Genetic Programming on Graphics Cards
Lessons
-
- Interpreting GP trees on the GPU is dominated by leafs
since there are lots of them and typically they require data transfers
across the graphics card.
- So having five operations (most of which are fast functions)
in the interpreter loop
imposes an overhead of
2.89 times rather than 5
(Mackey-Glass)
- To get the best of the GPU it needs to be give large chunks of
work to do.
- Aim for 1-10 seconds.
- More than about 16 seconds and Linux dies