Convergence Iterations - Conjugate Gradient CUDA

I am doing some research in the Conjugate Gradient solver and was wondering about number of iterations to converge on the GPU verses the CPU.  When I execute the CPU version of the preconditioned conjugate gradient solver, it takes more iterations to converge as compared to the GPU version.  The GPU always converges with less iterations - I use different CPUs and GPUs (one is NVIDIA Quadro fx5600 and the other is Tesla M2070) but same iteration count results.

_______________________________________________________________________________________________

Opensource Tools to create Multi GPU cluster across a network

 Hello Friends,

I have 4 Tesla C2050 cards and have hooked each of them to individual computers, i.e each forming a node. I want to run CUDA programs such that it uses the compute power of this cluster.

• Are there any Opensource tools that can be used to establish the cluster using these 4 nodes?
• How to write a CUDA program that is distributed across the GPU nodes?

_______________________________________________________________________________________________