Interesting project for sure. When I was working on machine learning project this summer I decided to use the GPU to do a lot of computation on a smallish data

Interesting project for sure. When I was working on machine learning project this summer I decided to use the GPU to do a lot of computation on a smallish dataset (100 MiB) with 250k records and the alogirthm was O(n^2) and at some points even O(n^3). I tried to use existing solutions (ViennaCL, etc) but alas nothing seemed to work fast or at all. In the end learning CUDA turned out to be quite easy and profiling with Nvidia's tools is very nice and for most problems it seems rolling your own solution is often the best as they can be so ridiculously optimized (100% bandwidth utilization on 33% thread occupancy)