There must be a relation between debug built (nvcc -g -G to debug with cuda-gdb) and GPU resource consumption?
I am using Ubuntu 16.04, toolkit 8.0 on a Titan V GPU and debugging some deep learning code in Tensorflow 1.4.0 with GPU back-end C++ code. As soon as I switch to build the code with debug information, it breaks with
cudaCheckError() too many resources requested for launch
noted that with release built, it does not have this issue although my code is rather close GPU limit. How could I manage to build and debug this code?