I have gone through the MKL documentation but find it very unclear which MKL routines can actually offload to Xeon Phi. Does the MKL include any sparse direct solver capable of offloading to Xeon phi, and if so, does anyone have experience with how it performs compared to regular Xeons?
Thanks in advance for your help :-)
actually, all MKL's functions ( including Sparse Direct Solver) can be offloaded to Xeon Phi. You can see the offload version of Pardiso into directory: < mklroot>\examples\mic_offload\solverc\source\. We have never check the performance of offloading version of Pardsio because of memory offload overhead may be significant.
All official performance data are published on the main MKL page - https://software.intel.com/en-us/intel-mkl/ . see "benchmarks/Intel Xeon Phi Coprocessor" bookmark. there are no such data for direct solver. The main reasons - there are many factors effect on Direct Solver performance and the RAM available in native mode <= 8 Gb