Simulation running well on K620 but instable on A100?

[maleadt]

@Ludovic_Dumoulin

It happens because each thread of index i add a float to an array at indices {i-2,…,i+2} {i−2,…,i+2}{i-2,…,i+2} , then the more a graphic card can run threads in parallel, the more there are conflicts between threads.

Yeah, don’t do that A quick fix is to add CUDA.@atomic in front of those additions. That will make it slower, though. A better solution is to compute interim sums, aggregate those across the block, and then perform atomic additions at the grid level; but that’s of course much more invasive. Depending on the exact characteristics, CUDA.@atomic might perform well enough for you.