Hi. I’m trying to compile TensorFlow from the master branch. Cuda8 has been installed successfully, but when building TensorFlow with cuda support, it fails with an error from tensorflow/third_party/gpus/cuda/include/host_config.h, saying that the gcc version cannot be higher than 5.3:
#ifGNUC > 5 || (GNUC == 5 && GNUC_MINOR > 3) #error – unsupported GNU version! gcc versions later than 5.3 are not supported!
I compared this file with host_config.h that comes from nvidia with the cuda8.0, and it’s the exact same file.
I saw on some thread on reddit (I think) that cuda8rc is supposed to support gcc6. It doesn’t seem like it from the files.
Any info would be appreciated.
Take a look at Table 1 in the CUDA_Installation_Guide_Linux.pdf that got installed on your machine when you installed CUDA 8RC. You’ll see that it lists the supported distros and supported gnu versions.
The highest gcc supported version listed there is 5.3.1
I have the same Problem, my default systemcompiler is the gcc-6, but i have found a workaround , i use as compileoptions -Xcompiler -std=c++98, this works, you can compile cuda programs with the gcc-6, my linux is openSUSE 13.3 Tumbleweed 64 Bit , the problem is, you must modify each Makefile manually the setting in the nvcc.profile NVCCFLAGS += -Xcompiler -std=c++98, does not work