Early CUDA programs had to conform to a flat, bulk parallel programming model. Programs had to perform a sequence of kernel launches, and for best performance each kernel had to expose enough parallelism to efficiently use the GPU. For applications consisting of “parallel for” loops the bulk parallel model is not too limiting, but some parallel patterns—such as nested parallelism—cannot be express
![Adaptive Parallel Computation with CUDA Dynamic Parallelism | NVIDIA Technical Blog](https://cdn-ak-scissors.b.st-hatena.com/image/square/dc2bcc7e5d246cc5998fd91c7277a7b8969b3506/height=288;version=1;width=512/https%3A%2F%2Fdeveloper-blogs.nvidia.com%2Fwp-content%2Fuploads%2F2014%2F04%2Fdynpar_thumb.png)