Unet-Segmentation
Unet-Segmentation copied to clipboard
lmb-freiburg
GCC version error for GPU installation
Hi,
I am trying to install GPU version on a backend server and facing this error with make -j install.
#error -- unsupported GNU version! gcc versions later than 6 are not supported!
Can you help me with the installation. Please let me know if you need any details.
The non GPU version works well on the same linux server.
Regards Rakesh Rakesh
Probably you must update your nVidia driver and/or cuda toolkit. I don't remember exactly which version worked with which gcc versions, but check the nVidia cuda installation page for details. CUDA 9 is the first version working with gcc6, If you cannot upgrade cuda toolkit you will have to install an older compiler e.g. gcc 4.8.
Dear Thorsten,
Thank you! I have now installed gcc 4.8 and that error is gone.
I am now facing error with the following message when I do make -j install :
nvcc fatal : Unsupported gpu architecture 'compute_20' nvcc fatal : Unsupported gpu architecture 'compute_20' CMake Error at cuda_compile_1_generated_contrastive_loss_layer.cu.o.Release.cmake:219 (message): Error generating /home/rakesh/caffe/x86_64/src/caffe/CMakeFiles/cuda_compile_1.dir/layers/./cuda_compile_1_generated_contrastive_loss_layer.cu.o
Can you please help sort this out.
Regards Rakesh
If you need CC 20 your GPU must be ancient. Remove the compute_20 entries (they are not in the guide, are they?) or further downgrade your cuda toolkit until you reach one that supports your GPU. Check the nVidia cuda site for the compatibility matrix. If you really need CC20 I cannot give any guarantees that caffe will build or behave as expected.
If you need CC 20 your GPU must be ancient. Remove the compute_20 entries (they are not in the guide, are they?) or further downgrade your cuda toolkit until you reach one that supports your GPU. Check the nVidia cuda site for the compatibility matrix. If you really need CC20 I cannot give any guarantees that caffe will build or behave as expected.
The compute_20 is actually still in the guide: " cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/home/unetuser/u-net -DUSE_OPENCV=OFF -DUSE_LEVELDB=OFF -DUSE_LMDB=OFF -DBUILD_python=OFF -DBUILD_python_layer=OFF -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN="20 30 35 50 60 61" -DCUDA_ARCH_PTX="35" .. "
