Hi all,
We have a simulation code written in C++ that solves Maxwell integral equations. MKL has been a great choice for us for several years and now we intend to compare the performance when the calculations are offloaded to GPUs. Our resources are enhanced with NVIDIA GPUs and we have implemented matrix factorization and solutions i.e. zgetrf + zgetrs since our data are complex double. Our measurements show that MKL on with 32 cores (CPU) is much faster than 2 NVIDIA Tesla T4 cards where each card has 2560 GPU cores. This is surprising and I wanted to first of all ask if anyone else has had such experience? Second, if this will be the case when Intel GPUs are used?
For GPU, we use native CUDA implementation of LAPACK which is called cuSolver.
Regards,
Dan
