The underlying mechanism is lazy loading. It has its own chapter in the CUDA C++ programming guide and has been discussed in a number of forum posts. Here is a recent example.
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| First kernels launch can't run in parallel? | 8 | 316 | October 28, 2025 | |
| Launching several kernels on one stream while another kernel running persistently in the background | 1 | 822 | October 8, 2016 | |
| Kernel launches/synchronization | 1 | 2176 | September 28, 2009 | |
| kernel launches in the same stream | 4 | 5397 | September 22, 2010 | |
| Is this strange behaviour with kernel concurrency? | 5 | 1096 | September 3, 2015 | |
| Does cudaLaunchHostFunc block work added to all streams? | 18 | 1956 | April 27, 2021 | |
| Kernels won't run in parallel | 3 | 1307 | May 8, 2013 | |
| Deadlock on cuda kernel launch | 12 | 858 | July 21, 2025 | |
| CUDA Parallel Kernels | 2 | 6988 | May 24, 2009 | |
| Running a kernel blocks the CPU? Is it possible to run it asynchronously? | 2 | 3659 | April 21, 2009 |