DBCSR-only GPU offload (OpenCL) on Intel PVC slower than CPU-only - expected?

[ganta.pra...@gmail.com]

Dear all,

I am running CP2K with the OpenCL backend on Intel Data Center GPU Max 1550 (Ponte Vecchio) nodes on SuperMUC-NG Phase 2, offloading only the DBCSR sparse matrix-matrix multiplication library to the GPU while the rest of the workload remains on the CPU.

Benchmark results (below) show that the GPU-offloaded runs are consistently slower than the CPU-only runs. I assume that because only DBCSR is offloaded, the host-device data transfer and synchronization overhead outweighs the compute gain from the GPU 

Has this behavior of GPU offload underperforming CPU-only execution when only DBCSR is accelerated been reported before? I'd also appreciate pointers to published OpenCL benchmark results for DBCSR/CP2K on Intel GPUs.

Below are the results 

Have a nice day.

Thanks and Regards,

Prasanth.