Simple or double precision GPU?

Getting Started
#1

Hi,

We’ll looking to buy GPUs for AI model trainings from Geant4/Gate generated data, so that the AI model would allow fast and decently accurate spatial radiation distribution on various source/shielding configurations by just a few real measure points.

Then we realized we could also have GPU acceleration of the simulations, and GGEMS came as a GPU-aware Geant4/Gate alternative. However, the price difference between simple and double precision GPU card is very large. Simple precision is enough for AI, but we don’t know if that’s also the case for GGEMS.

Can you enlighten me? Thanks,
Regards,

Yannick

#2

Hello Yannick,

In GGEMS, single precision is sufficient and, of course, faster than double precision. In most cases, single precision is adequate in GGEMS. For instance, applications that require double precision in GGEMS are dosimetry simulations, where the energies in the voxels are often incremented, and therefore “+=” operation needs to be precise. However, GGEMS also works very well on CPU, so don’t hesitate to check if single precision is sufficient in your case.

Kind regards
Didier

#3

Hello Didier,

Thanks for your answer. It seems that simple precision is enough in our case.

Cheers,

Yannick