Hi all, I agree with Andreas: Intel did not put much effort on the OpenCL driver for the Phi. Moreover the MPSS deamon (controlling the Phi) was running at full speed on one core.
The Phi is probably outperformed by their Iris-Pro with Beignet drivers on a laptop ! I published some benchmarks on the Phi vs Tesla or AMD GPUs: http://arxiv.org/pdf/1412.6367v1.pdf The overall conclusions: In average the Phi has the performance of a couple of high-end Xeons (bought at the same time). The Phi exhibits performances of a GPU when memory is perfectly aligned but the performances are 10x slower than a CPU when the memory pattern is either irregular or misaligned (no demonstrated into publication). Kernel often crash on the Phi when they perfectly run everywhere else ! Cheers, Jerome _______________________________________________ PyOpenCL mailing list PyOpenCL@tiker.net http://lists.tiker.net/listinfo/pyopencl
