TL;DR Guide for 2017 Macbook Pro: connect to GTX 1080 Ti, install CUDA and CUDNN, build tensorflow 1.3 with gpu support.


I spent two days running into problems and finally resolved to get my tensorflow powered by GTX 1080 Ti. Cheers! Thanks to all the online posts, most problems I got were already answered. There were a few issues raised because of versions of bazel and libraries involved. So I would clarify as much as possible on those.


Equipments I had:

  • 2017 MacBook Pro, 13-inch without touch bar. OS build number 16F2073
  • MSI GTX 1080 Ti.
  • Akitio Node Thunderbolt 3 External, Bought from BH Photo Video.

Environment I used:

  • homebrew 1.3.4
  • anaconda + python 3.6
  • I created a env in anaconda use conda create --name [env_name] python=3.5 numpy scipy matplotlib theano keras ipython jupyter and pip 9.0.1 comes with py35.


I relied on these great tutorials to set things up.

Bash Profile

Throughout the process I added a bunch of paths to my bash profile. Here is a summary of them:

# =================================
# Set up eGPU driver + tensorflow
# =================================
# Step 1.
# so that all the CUDA binaries are available to you on the command line:
export PATH=/Developer/NVIDIA/CUDA-8.0/bin${PATH:+:${PATH}}
# Step 2.
# CUDNN Installation
export DYLD_LIBRARY_PATH="/usr/local/cuda/lib":$DYLD_LIBRARY_PATH

# Step 3. 
# Set up env to build tensorflow
# copied from

export CUDA_HOME=/usr/local/cuda
export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:/usr/local/cuda/extras/CUPTI/lib
# ==================================


  1. Install CUDA from Nvidia and follow its official instructions. Or, as suggested by this post, install with brew tap caskroom/drivers & brew cask install cuda. I installed from Nvidia, chose CUDA 8.0.61 and patched it to 8.0.62.

  2. Connect Graphic Card to Akitio Node and to laptop, screw-driver needed. Youtube Video

  3. Disable SIP (System Integrity Protection). Tutorial

  4. Change OS Build Version (required for Macbook with build 16F2073). Discussion on this. In short, open the file /System/Library/CoreServices/SystemVersion.plist and change the build number from 16F2073 to 16F73.

  5. Download the script and execute it.

    sudo ./ -url
  6. After restart, I upgrade CUDA to 8.0.90

  7. Downgrade Xcode to 8.2 and corresponding command line tools. You could download them from Apple Developer website. Newer versions of clang would give you error when compiling CUDA samples.

    Verify by clang —version and pkgutil Expecting Apple LLVM version 8.0.0 (clang-800.0.42.1) and version:

  8. Add CUDA binaries to path

    export PATH=/Developer/NVIDIA/CUDA-8.0/bin${PATH:+:${PATH}}
  9. Verify CUDA installation by running deviceQuery

    cd /usr/local/cuda/samples
    sudo make -C 1_Utilities/deviceQuery

    should detect the device and yield CUDA Capability Major/Minor version number: 6.1 ( I have a Geforce GTX 1080 Ti).

  10. Install cuDNN v6.0 for Mac OS

    • Download and unzip cuDNN 6.0 for MacOS from NVIDIA

    • Move the cuDNN libraries to cuda:

     sudo mv -v cuda/lib/libcudnn* /usr/local/cuda/lib
     sudo mv -v cuda/include/cudnn.h /usr/local/cuda/include
     sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib/libcud
    • Add to path
     export DYLD_LIBRARY_PATH="/usr/local/cuda/lib":$DYLD_LIBRARY_PATH
    • Verify the installation by echo -e '#include"cudnn.h"\n void main(){}' | nvcc -x c - -o /dev/null -I/usr/local/cuda/include -L/usr/local/cuda/lib -lcudnn.

    • I got a few warning but no error.

  11. Installing tensorflow from Source, following the official doc except the following:

    • Bazel version: The most recent r1.3 branch of tensorflow asks for bazel version 0.5.4. I got xxx file not built error when building with bazel 0.5.4. Therefore, I cd tensorflow & git checkout b46340f and use bazel 0.4.5 that comes with conda to build.

    • Set the environment flag

       export CUDA_HOME=/usr/local/cuda
       export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:/usr/local/cuda/extras/CUPTI/lib
    • After checkout, comment out the BUILD file requiring OpenMP: open tensorflow/third_party/gpus/cuda/BUILD.tpl and comment out # linkopts = [“-lgomp”]

    • Checksum mismatch discussion on Github. Workaround: If it happens, comment out the checksum line in tensorflow/workspace.bzl that says sha256=repo_ctx_attr.sha256.

    • run the configuration, opt in for CUDA support, and substitute TF_CUDA_COMPUTE_CAPABILITIES with your output of deviceQuery.

    • bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package --verbose_failures --action_env PATH --action_env LD_LIBRARY_PATH --action_env DYLD_LIBRARY_PATH

      Took about 50 minutes and huge RAM consumption. Expecting no error.

    • Build the pip package bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

    • Install the pip package sudo pip install /tmp/tensorflow_pkg/tensorflow-1.3.0-.whl

  12. ALL DONE!

    • Test your installment by running the cutest MNIST toy

    • See, in your terminal, that tensorflow is running!

        2017-09-26 21:34:32.053838: I tensorflow/core/common_runtime/gpu/] DMA: 0
        2017-09-26 21:34:32.053843: I tensorflow/core/common_runtime/gpu/] 0:   Y
        2017-09-26 21:34:32.053851: I tensorflow/core/common_runtime/gpu/] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:85:00.0)