I find it interesting that NVIDIA has been decreasing the number of CUDA cores per streaming multiprocessor (SM) in each successive architecture, while increasi

I find it interesting that NVIDIA has been decreasing the number of CUDA cores per streaming multiprocessor (SM) in each successive architecture, while increasing the number of streaming multiprocessors. It makes each SM look more like a CPU core with a vector unit. I wonder what the reason it might be - it doesn't really change the programming model, it just requires you to tune the size of thread blocks versus number of thread blocks differently.