The Tesla is essentially a graphics card with no output. So essentially it specialises at SIMD (Single instruction multiple data) So think data parallel tasks.
Xeon Phi will be superior in logical/arithmetic calculations, whereas GPGPU will be good in FP calculations.
The Tesla needs to have all of it's data loaded and unloaded before it can do anything. So there will be noticeable latency in your algorithm
You can probably google as much as I can. Look at benchmarks that match your application.