Thanks, I came to ask that question also. So is there any reason that this can't be done using OpenCL instead?
It's just too slow to move it from main to video memory and back.
Also OpenCL works best when working on large pieces of data at once, with High frequency trading often you want to work on very small pieces, 50 bytes, and process them as fast as possible. Batching up the data and working on it in a SIMD type fashion just takes too long:(