Why does a CUDA program work with CMake “FIND_PACKAGE”, but not with “LANGUAGES CUDA”?

Note: previous title was "How can I compile a CUDA program with CMake, without using FIND_PACKAGE?". Edited to make it clearer what the key issue is.

When using CMake and CUDA in the suggested way, without FIND_PACKAGE, it compiles, but when I run the resulting binary, I get the following error:

CUDA error: CUDA driver version is insufficient for CUDA runtime version

However, if I manually compile with nvcc, or if I use FIND_PACKAGE in cmake, it works fine. The CUDA examples compile and run also. I understand that FIND_PACKAGE(CUDA) is deprecated, because cuda is now a first class cmake language. What is the difference between the two different approaches, and how do I get it to work without using find package?

Details:

main.cu:

#include <iostream>

__global__ void helloGPU() {
    printf("Hi globe\n");
}

int main() {
    helloGPU<<<1,2>>>();
    cudaDeviceSynchronize();
    cudaError_t error = cudaGetLastError();
    if(error != cudaSuccess)
    {
        printf("CUDA error: %s\n", cudaGetErrorString(error));
        exit(-1);
    }
}

CMakeLists.txt that works:

cmake_minimum_required (VERSION 3.10)
find_package(CUDA REQUIRED)
cuda_add_executable(simple main.cu)

CMakeLists.txt that is broken:

cmake_minimum_required (VERSION 3.10)
project(simplecuda LANGUAGES CXX CUDA)
add_executable(simple main.cu)

System:

Mac OS High Sierra 10.13.6
Cuda driver version: 410.130
GPU Driver Version: 387.10.10.10.40.105
nvcc version: 10.0.130
clang++ version: 9.0.0

Edit: why the current driver versions

Some comments were suggesting different driver versions. Here is why I have the listed versions:

With the default driver installed by the CUDA installer, I got the following error when trying to run the CUDA samples:

CUDA driver version is insufficient for CUDA runtime version

Sites like this persuaded me that trying a different driver version could help: (https://devtalk.nvidia.com/default/topic/1027922/cuda-setup-and-installation/-solved-code-35-cudaerrorinsufficientdriver-error-on-mac-version-10-13-2-17c88-with-nvidia-geforce-gt-/).

After some unsuccessful trial and error, I came across this script: https://www.tonymacx86.com/threads/nvidia-update-simple-way-to-install-nvidia-web-drivers.244987/

This automatically searches for and installs a driver, which is how I got 387.10.10.10.40.105.

With this version, for the first time, the CUDA samples work fine, manually compiling source with nvcc works, and using FIND_PACKAGE works. I am relatively certain that my driver versions are now correct. When I use NVIDIA Driver Manager preference pane to check for new versions, it says I have the latest, and according to http://www.macvidcards.com/drivers.html, I have the correct driver version.

Edit: digging into build commands

I dug further into what CMake is doing by using make VERBOSE=1. I removed as many lines and flags as I could to still get the same behaviour, and simplified paths, tmp directories etc. Here is a minimal example of the commands to produce a binary that gives an error, taken from the CMake with LANGUAGES CXX CUDA (if I run these in the same folder as main.cu, then run ./simple, it gives the CUDA error above):

/Developer/NVIDIA/CUDA-10.0/bin/nvcc  -c main.cu -o main.cu.o
/Developer/NVIDIA/CUDA-10.0/bin/nvcc  -dlink main.cu.o -o cmake_device_link.o
/usr/bin/clang++  main.cu.o cmake_device_link.o -o simple  -L"/Developer/NVIDIA/CUDA-10.0/lib" "/Developer/NVIDIA/CUDA-10.0/lib/libcudart_static.a"

Here is a similar minimal version of the commands from the CMake with FIND_PACKAGE(CUDA):

/Developer/NVIDIA/CUDA-10.0/bin/nvcc main.cu -c -o main.cu.o -I/Developer/NVIDIA/CUDA-10.0/include
/usr/bin/clang++   main.cu.o  -o simple /Developer/NVIDIA/CUDA-10.0/lib/libcudart_static.a -Wl,-rpath,/usr/local/cuda/lib

The critical part seems to be the very last flag, -Wl,-rpath,/usr/local/cuda/lib. I can get the broken CMake version to work by either setting an environment variable:

export LD_LIBRARY_PATH=/usr/local/cuda/lib

or by updating the CMake file as follows:

SET_TARGET_PROPERTIES(simple PROPERTIES LINK_FLAGS -Wl,-rpath,/usr/local/cuda/lib)

Is something broken on my system that I need to do this? Or broken in CMake's cuda compatibility on Mac?

Note: there is something about Apple and rpaths mentioned in the official blog: https://devblogs.nvidia.com/building-cuda-applications-cmake/

But if I add this section to CMakeLists.txt, it does not help:

if(APPLE)
  # We need to add the path to the driver (libcuda.dylib) as an rpath, 
  # so that the static cuda runtime can find it at runtime.
  set_property(TARGET simple
               PROPERTY
               BUILD_RPATH ${CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES})
endif()