Previous | Next --- Slide 68 of 74
Back to Lecture Thumbnails

How do context switches work in CUDA? It seems like we generally want to avoid them to maximize arith. intensity


I thought that warps in a thread block could be scheduled across different SMs like in the previous fictitious example with 4 warps but 8 warps were needed for the thread block


^Oh, I realize that those were subcores in an SM

Please log in to leave a comment.