OSX Sierra + Titan Xp eGPU. Tensorflow 1.2.1. + CUDA 8 + cuDNN 6

OSX Sierra + Titan Xp eGPU. Tensorflow 1.2.1. + CUDA 8 + cuDNN 6.md

OSX Sierra + Titan Xp eGPU. Tensorflow 1.2.1. + CUDA 8 + cuDNN 6

This gist summarises the tensorflow related steps I took to get the above combo working.

When its all tested in c++ I'll update the instructions fully (and add anything I forgot below).

But in summary ( with the current master dd06643cf098ed362212ce0f76ee746951466e81 ):

I have uploaded the pip wheel which I believe should work if you have the same setup but no promises (built for compute capability 3.5, 5.2, 6.0 and named tensorflow-gpu). Install with (not sure dropbox allows this direct linking):

pip install http://dl.dropboxusercontent.com/s/reo3pkz6dn33u8k/tensorflow_gpu-1.2.1-cp27-cp27m-macosx_10_11_x86_64.whl

Or download then install

https://www.dropbox.com/s/reo3pkz6dn33u8k/tensorflow_gpu-1.2.1-cp27-cp27m-macosx_10_11_x86_64.whl?dl=1

pip install tensorflow_gpu-1.2.1-cp27-cp27m-macosx_10_11_x86_64.whl

• CUDA 8.0.61
  • export PATH=/usr/local/cuda/bin:$PATH
  • export LD_LIBRARY_PATH=/usr/local/cuda/lib:$LD_LIBRARY_PATH
  • export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:$DYLD_LIBRARY_PATH
• cuDNN 6.0.21
• NVidia Web Driver 378.05.05.15f01
• LLVM but older command line tools to keep nvcc happy (xcode-select version 2347)
• Add libgomp from GCC... to the library paths (brew install gcc)
  • export LD_LIBRARY_PATH=/usr/local/Cellar/gcc/7.1.0/lib/gcc/7:$LD_LIBRARY_PATH
  • export DYLD_LIBRARY_PATH=/usr/local/Cellar/gcc/7.1.0/lib/gcc/7:$DYLD_LIBRARY_PATH
• Normal configure steps (set your CUDA compute capability)
• Export your library paths to the bazel build with --action_env LD_LIBRARY_PATH="$LD_LIBRARY_PATH". Not sure if this is necessary but was needed at one point, Ill try when I next build and drop it from here if unecessary. So for example if you also want libtensorflow.so :
  • bazel build --config=opt --config=cuda --show_result 100 //tensorflow/tools/pip_package:build_pip_package //tensorflow/python/tools:freeze_graph //tensorflow/python/tools:optimize_for_inference //tensorflow:libtensorflow.so //tensorflow/cc:tutorials_example_trainer --action_env LD_LIBRARY_PATH="$LD_LIBRARY_PATH"
• Build pip package if you want to use python
  • bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
• Install
  • sudo pip install /tmp/tensorflow_pkg/<OUTPUT_WHEEL>.whl

Let me know if you have any problems as im sure I forgot some steps

Hey, I will definitely try out your guide! Thanks!

Did you do anything else to setup your eGPU, like using this tool https://egpu.io/forums/mac-setup/imac-egpu-simply-for-3d-accelleration/?

Did you do any benchmarks? Is TF playing nicely or is there a significant performance loss compared to using this GPU with Windows?

Thank you so much!
Dennis

Can you please build the same wheel for Python 3.5?

Hi!

When training neural networks, how 'hot' the computer gets? I am not referring to the eGPU but the computer itself.

I have got a macbook pro late 2013 and I am not sure if I should add an eGPU for machine learning purposes.

Thanks!

Miguel

What eGPU did you use?

Hi @danbarnes333,

Thanks for this tutorial!

I have followed your instructions as well as I could, but I'm running into the ld: library not found for -lgomp error. I have added the locations of libgomp.dylib and libgomp.a to both the DYLD_LIBRARY_PATH and the LD_LIBRARY_PATH but to no avail. What am I missing?

$ gcc -v
Configured with: --prefix=/Applications/Xcode_8.2.1.app/Contents/Developer/usr --with-gxx-include-dir=/usr/include/c++/4.2.1
Apple LLVM version 8.0.0 (clang-800.0.42.1)
Target: x86_64-apple-darwin16.7.0
Thread model: posix
InstalledDir: /Applications/Xcode_8.2.1.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin

$ /opt/local/bin/gcc-mp-7 -v
Using built-in specs.
COLLECT_GCC=/opt/local/bin/gcc-mp-7
COLLECT_LTO_WRAPPER=/opt/local/libexec/gcc/x86_64-apple-darwin16/7.2.0/lto-wrapper
Target: x86_64-apple-darwin16
Configured with: /opt/local/var/macports/build/_opt_bblocal_var_buildworker_ports_build_ports_lang_gcc7/gcc7/work/gcc-7.2.0/configure --prefix=/opt/local --build=x86_64-apple-darwin16 --enable-languages=c,c++,objc,obj-c++,lto,fortran --libdir=/opt/local/lib/gcc7 --includedir=/opt/local/include/gcc7 --infodir=/opt/local/share/info --mandir=/opt/local/share/man --datarootdir=/opt/local/share/gcc-7 --with-local-prefix=/opt/local --with-system-zlib --disable-nls --program-suffix=-mp-7 --with-gxx-include-dir=/opt/local/include/gcc7/c++/ --with-gmp=/opt/local --with-mpfr=/opt/local --with-mpc=/opt/local --with-isl=/opt/local --enable-stage1-checking --disable-multilib --enable-lto --enable-libstdcxx-time --with-build-config=bootstrap-debug --with-as=/opt/local/bin/as --with-ld=/opt/local/bin/ld --with-ar=/opt/local/bin/ar --with-bugurl=https://trac.macports.org/newticket --disable-tls --with-pkgversion='MacPorts gcc7 7.2.0_0'
Thread model: posix
gcc version 7.2.0 (MacPorts gcc7 7.2.0_0)

I was able to solve/circumvent the problem by applying this patch to ./third_party/gpus/cuda/BUILD.tpl:

--- a/third_party/gpus/cuda/BUILD.tpl
+++ b/third_party/gpus/cuda/BUILD.tpl
@@ -109,7 +109,7 @@ cc_library(
         ".",
         "cuda/include",
     ],
-    linkopts = ["-lgomp"],
+    linkopts = ["-L/opt/local/lib/gcc7/libgomp.dylib"],
     linkstatic = 1,
     visibility = ["//visibility:public"],
 )

see alse https://metakermit.com/2017/compiling-tensorflow-with-gpu-support-on-a-macbook-pro/