Comment by esychology

I think performance is not the only issue for scaling to larger grids. CUDA Convolution implementation already utilizes coalescing to improve performance. The main bottleneck is that in larger grids, cells are further apart, and it takes more steps for them to be able to communicate.