search

simone-silvestri / CudaNavierStokes

4 stars 1 forks source link

Reduce the memory footprint on the GPU #13

Open simone-silvestri opened 3 years ago

simone-silvestri commented 3 years ago

Try to remove (or change the way we do) the velocity derivative tensor such to not store the whole tensor.

For the moment there are 41 fields living on a GPU

Therefore, if the GPU has 12GB of memory, the allowed number of points (in double precision) are mx my mz = (12 1024 1024 1024) / (41 8) = 39.2 M

simone-silvestri commented 3 years ago

Now with the low storgae Runge Kutta, only 31 fields are used.