Regarding GPU requirement for large domain and improved simulation time

[MOLOY DAS]

Dear All,

As per my limited knowledge on working with gprMax software, I found that:

1. GPU RAM is related to the size of the domain that may be accommodated. 

2. The simulation in it/sec is related to the number of CUDA cores.

We are presently using NVIDIA RTX A2000 and A4000 series getting speed of 85 to 130 it/sec for 1.8 X 0.24 X 0.600 domain. We have tried with single A2000, A4000+ A2000, A4000+A4000 and the simulation time is improved from 2 hr 50 mins to 1 hr 4 mins at best. GPU parallelization is conducted successfully to achieve this by using this command:

"mpiexec -n 3 python -m gprMax file path -n 157 --mpi-no-spawn -gpu 0 1" with CPU working as master and the two GPUs acting as workers.

In near future we want around 4 meter homogenous domain size to be completed in 45 minutes approximately so that around 1000 B scans can be generated in a month with the system continuously working.

Please provide your suggestions on the GPU requirement for such task.

Thanking You,

Moloy Narayan Das,

Senior Scientist,

CSIR-CMERI,

India