Hardware Benchmarking Topic HPC (GPU) vs GPU Solver

From Tuflow
Jump to: navigation, search

Page Under Construction

Introduction

As discussed elsewhere (????) TUFLOW HPC can run on either CPU or GPU hardware. When running on GPU devices TUFLOW HPC uses all computational (CUDA) cores on the GPU. It is also possible to run a single HPC simulation across multiple GPU devices if these are available in the computer. When using multiple GPU devices, at each timestep there needs to be a transfer of information between GPU devices. This adds an overhead and means that the scaling is never perfectly linear. For example, running on two GPU devices is not twice as fast as running on a single GPU.
This page discusses the speed benefits when running a range of models across different numbers of GPU devices. For a discussion on scalability of TUFLOW HPC on different numbers of CPU cores please refer to HPC (CPU) Scaling

Results

???? Charts of scalability or runtime with varying numbers of GPU devicesare required. It would be good to have this for a couple of cell sizes / models. E.g. FMA 2 at 10m, 5m and maybe finer at 1,2,4, GPUs. The take home message is that as the size of the model increases the better the scalign across multiple GPU devices. I'm not sure where we are going to find these multiple GPU machines...