This repository contains mini-apps that can be used for experimenting with new implementations for the core algorithm's in the Spectral Element Library in Fortran.
These results show the effective bandwidth (in GB/s) for the various 2-D divergence kernel implementations on Fluid Numerics' "Noether" platform. This system is equipped with 4x AMD MI210 GPUs and two AMD EPYC 7313 16-Core Processors, giving a 8 vCPU/GPU ratio. All benchmarks are shown for 1 MI210 GPU; for CPU-only kernels, we use 8 threads for our do concurrent loops. The environment consists of the Ubuntu 22.04 operating system, with gfortran 12.3.0, and ROCm 6.2.1 .
These results show the effective bandwidth (in GB/s) for the various 3-D divergence kernel implementations on Fluid Numerics' "Noether" platform. This system is equipped with 4x AMD MI210 GPUs and two AMD EPYC 7313 16-Core Processors, giving a 8 vCPU/GPU ratio. All benchmarks are shown for 1 MI210 GPU; for CPU-only kernels, we use 8 threads for our do concurrent loops. The environment consists of the Ubuntu 22.04 operating system, with gfortran 12.3.0, and ROCm 6.2.1