So I think this the reason why the comment in Assignment 2 points out the 256 threads per block is a healthy number even though the maximum block dimension of GTX 480 is 1024x1024x64. This gives the GPU chance to schedule thread blocks in order to hide memory latency, which maximum the parallelism. The number 256 may vary due to different types of work, however, the basic idea should still hold.
So I think this the reason why the comment in Assignment 2 points out the 256 threads per block is a healthy number even though the maximum block dimension of GTX 480 is 1024x1024x64. This gives the GPU chance to schedule thread blocks in order to hide memory latency, which maximum the parallelism. The number 256 may vary due to different types of work, however, the basic idea should still hold.
@BigFish Perhaps you will find this an interesting read. NVidia best practices guide.