You can launch much more logical threads than the available physical threads. The GPU scheduler will automatically dispatch the work to the SMs.