SYCL Graph in GROMACS: preliminary benchmarks

Preliminary benchmarks showing the performance effect of SYCL Graph in GROMACS. Presented performance is the median of five runs.

Benchmarks on Intel A770 use the PME variant as more representative of typical atomistic simulations and the one where the effect of SYCL Graphs is expected to be more pronounced due to larger number of kernels per step. Due to current limitations of FFT library / SYCL Graph interoperability, the benchmarks on NVIDIA V100 use RF variant of the benchmark systems.

• GROMACS commit: e55d44a76346b8864096ab76e47a735a9faa896b
• Benchmark systems: https://zenodo.org/records/11234002, grappa-1.5k-6.1M_rc0.9.tar.gz
• GROMACS flags (RF): -nb gpu -update gpu -bonded gpu -ntmpi 1 -pin on -nsteps 52000 -resetstep 2000 -nobackup -noconfout -dlb no -nstlist 200
• GROMACS flags (PME): -nb gpu -update gpu -bonded gpu -pme gpu -notunepme -ntmpi 1 -pin on -nsteps 52000 -resetstep 2000 -nobackup -noconfout -dlb no -nstlist 200
• Intel oneAPI version: 2024.2
• CUDA version: 12.2