GPU required specifications about mcx

[Eidantz]

I saw the list: http://mcx.space/gpubench/
Sorry for bothering you
Eidan.

[fangq]

The estimation of the total memory needed by an MCX simulation is straightforward, from your script, you requested a volume of 500x500x1500 voxels, and a total of 10 time-gates, each voxel uses a 4-byte single-precision number to store fluence. if you convert this to bytes, you need 50050015004/1024/102410 = 14 GB global memory. In addition, MCX need twice of that size internally to improve accuracy, so, your simulation requires 14*2 = 28GB GPU memory.

if you look at all the GPUs on the market, you can almost never find a GPU with such amount of memory, except the super expensive v100 GPU. So, your simulation is just not realistic.

Even you can find a GPU with such a large memory, to produce a fluence map that every voxel has decent SNR will require a huge amount of photons. The runtime will also be unrealistic. Think about this, if you run a simulation on a grid with 100x100x100 voxels, if you refine the grid to 200x200x200, you will need 64 times more run-times to produce a simulation with comparable SNR - 8 fold comes from the increase number of voxels, and another 8-fold comes from launching 8x more photons to produce the same SNR.

My suggestion is always: start with a small domain, run a small number of photons, and see if your solution is acceptable, then gradually increase the photon number or domain size. Do not go the other way around.