We currently run SL 6 on our compute engine.  It is multinode, 
internally connected by an Infiniband QDR 12x compute fabric (with plans 
to go to EDR 12X before NDR becomes production) and an 802.3ae (10 
Gbit/s) control fabric, using X86-64 CPUs and Nvidia Tesla GPUs at each 
node currently under Nvidia CUDA 6.  We are considering incorporating 
Intel Xeon Phi units to supplement and/or replace the GPUs.

We are (somewhat) familiar with CUDA and GPU issues, and the problems 
associated with porting/programming/designing applications for CUDA.  
Supposedly the Xeon Phi eliminates much of the syntactical complexity of 
a CUDA application, in so far as any multithreaded parallel (concurrent) 
application can be simplified but still exceed that of a simple 
uniprocessor software design.

Does anyone have any experience with the Xeon Phi in  real production 
research environments as either a development or porting platform, 
particularly in comparison with real experience on the Nvidia Tesla GPU 
CUDA platform and environment, under SL 6?  How much support is there 
for the Intel Xeon Phi under SL? Correspondence off-list is fine.

Yasha Karant