FAQ

Can I use my Fermi GPU with compute capability 2.1 to run caffe_unet?

In principle yes, but the pre-built binaries require at least compute capability 3. You can build from source, though. When using Ubuntu 16.04 with g++ 5.4 it works out-of-the box using CUDA 8.0.61 and corresponding cuDNN. Things changed with Ubuntu 18.04 and g++ 7, because CUDA 8 and gcc 7 are incompatible. You have to explicitly install an older g++ (e.g. 5.4) and manually enable c++11 features by adding the line

 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")

to CMakeLists.txt. This is required because many libraries in Ubuntu 18.04 are built with gcc 7 with c++11 features enabled by default. Probably you should think about buying a more recent GPU, a GTX1080 is quite cheap, and allows you to use the recent CUDA Toolkit without these hassles!