diff --git a/configure.py b/configure.py
index 3365595..073bcca 100644
--- a/configure.py
+++ b/configure.py
@@ -518,6 +518,21 @@
       no_reply=no_reply)
 
 
+def set_tf_download_clang(environ_cp):
+  """Set TF_DOWNLOAD_CLANG action_env."""
+  question = 'Do you want to download a fresh release of clang? (Experimental)'
+  yes_reply = 'Clang will be downloaded and used to compile tensorflow.'
+  no_reply = 'Clang will not be downloaded.'
+  set_action_env_var(
+      environ_cp,
+      'TF_DOWNLOAD_CLANG',
+      None,
+      False,
+      question=question,
+      yes_reply=yes_reply,
+      no_reply=no_reply)
+
+
 def get_from_env_or_user_or_default(environ_cp, var_name, ask_for_var,
                                     var_default):
   """Get var_name either from env, or user or default.
@@ -1301,8 +1316,19 @@
 
     set_tf_cuda_clang(environ_cp)
     if environ_cp.get('TF_CUDA_CLANG') == '1':
-      # Set up which clang we should use as the cuda / host compiler.
-      set_clang_cuda_compiler_path(environ_cp)
+      if not is_windows():
+        # Ask if we want to download clang release while building.
+        set_tf_download_clang(environ_cp)
+      else:
+        # We use bazel's generated crosstool on Windows and there is no
+        # way to provide downloaded toolchain for that yet.
+        # TODO(ibiryukov): Investigate using clang as a cuda compiler on
+        # Windows.
+        environ_cp['TF_DOWNLOAD_CLANG'] = '0'
+
+      if environ_cp.get('TF_DOWNLOAD_CLANG') != '1':
+        # Set up which clang we should use as the cuda / host compiler.
+        set_clang_cuda_compiler_path(environ_cp)
     else:
       # Set up which gcc nvcc should use as the host compiler
       # No need to set this on Windows
diff --git a/third_party/gpus/cuda_configure.bzl b/third_party/gpus/cuda_configure.bzl
index 31a4bfa..5f1c42d 100644
--- a/third_party/gpus/cuda_configure.bzl
+++ b/third_party/gpus/cuda_configure.bzl
@@ -8,6 +8,9 @@
   * `TF_CUDA_CLANG`: Whether to use clang as a cuda compiler.
   * `CLANG_CUDA_COMPILER_PATH`: The clang compiler path that will be used for
     both host and device code compilation if TF_CUDA_CLANG is 1.
+  * `TF_DOWNLOAD_CLANG`: Whether to download a recent release of clang
+    compiler and use it to build tensorflow. When this option is set
+    CLANG_CUDA_COMPILER_PATH is ignored.
   * `CUDA_TOOLKIT_PATH`: The path to the CUDA toolkit. Default is
     `/usr/local/cuda`.
   * `TF_CUDA_VERSION`: The version of the CUDA toolkit. If this is blank, then
@@ -27,6 +30,7 @@
 _CUDNN_INSTALL_PATH = "CUDNN_INSTALL_PATH"
 _TF_CUDA_COMPUTE_CAPABILITIES = "TF_CUDA_COMPUTE_CAPABILITIES"
 _TF_CUDA_CONFIG_REPO = "TF_CUDA_CONFIG_REPO"
+_TF_DOWNLOAD_CLANG = "TF_DOWNLOAD_CLANG"
 
 _DEFAULT_CUDA_VERSION = ""
 _DEFAULT_CUDNN_VERSION = ""
@@ -34,6 +38,7 @@
 _DEFAULT_CUDNN_INSTALL_PATH = "/usr/local/cuda"
 _DEFAULT_CUDA_COMPUTE_CAPABILITIES = ["3.5", "5.2"]
 
+load(":download_clang.bzl", "download_clang")
 
 # TODO(dzc): Once these functions have been factored out of Bazel's
 # cc_configure.bzl, load them from @bazel_tools instead.
@@ -48,6 +53,8 @@
   if _use_cuda_clang(repository_ctx):
     target_cc_name = "clang"
     cc_path_envvar = _CLANG_CUDA_COMPILER_PATH
+    if _flag_enabled(repository_ctx, _TF_DOWNLOAD_CLANG):
+      return "extra_tools/bin/clang"
   else:
     target_cc_name = "gcc"
     cc_path_envvar = _GCC_HOST_COMPILER_PATH
@@ -80,6 +87,23 @@
     path = path[:-_OSX_FRAMEWORK_SUFFIX_LEN].strip()
   return path
 
+
+def _normalize_include_path(repository_ctx, path):
+  """Normalizes include paths before writing them to the crosstool.
+
+  If path points inside the 'crosstool' folder of the repository, a relative
+  path is returned.
+  If path points outside the 'crosstool' folder, an absolute path is returned.
+  """
+  path = str(repository_ctx.path(path))
+  crosstool_folder = str(repository_ctx.path(".").get_child('crosstool'))
+
+  if path.startswith(crosstool_folder):
+    # We drop the path to "$REPO/crosstool" and a trailing path separator.
+    return path[len(crosstool_folder)+1:]
+  return path
+
+
 def _get_cxx_inc_directories_impl(repository_ctx, cc, lang_is_cpp):
   """Compute the list of default C or C++ include directories."""
   if lang_is_cpp:
@@ -106,8 +130,11 @@
   else:
     inc_dirs = result.stderr[index1 + 1:index2].strip()
 
-  return [str(repository_ctx.path(_cxx_inc_convert(p)))
-          for p in inc_dirs.split("\n")]
+  return [
+      _normalize_include_path(repository_ctx, _cxx_inc_convert(p))
+      for p in inc_dirs.split("\n")
+  ]
+
 
 def get_cxx_inc_directories(repository_ctx, cc):
   """Compute the list of default C and C++ include directories."""
@@ -884,12 +911,14 @@
     result = find_result.stdout
   return result
 
+def _flag_enabled(repository_ctx, flag_name):
+  if flag_name in repository_ctx.os.environ:
+    value = repository_ctx.os.environ[flag_name].strip()
+    return value == "1"
+  return False
 
 def _use_cuda_clang(repository_ctx):
-  if "TF_CUDA_CLANG" in repository_ctx.os.environ:
-    enable_cuda = repository_ctx.os.environ["TF_CUDA_CLANG"].strip()
-    return enable_cuda == "1"
-  return False
+  return _flag_enabled(repository_ctx, "TF_CUDA_CLANG")
 
 def _compute_cuda_extra_copts(repository_ctx, compute_capabilities):
   if _use_cuda_clang(repository_ctx):
@@ -970,15 +999,25 @@
            "%{cuda_headers}": ('":cuda-include",\n' +
                                '        ":cudnn-include",')
        })
+
+  is_cuda_clang = _use_cuda_clang(repository_ctx)
+
+  should_download_clang = is_cuda_clang and _flag_enabled(
+      repository_ctx, _TF_DOWNLOAD_CLANG)
+  if should_download_clang:
+    download_clang(repository_ctx, "crosstool/extra_tools")
+
   # Set up crosstool/
   cc = find_cc(repository_ctx)
-  host_compiler_includes = _host_compiler_includes(repository_ctx, cc)
+  cc_fullpath = cc if not should_download_clang else "crosstool/" + cc
+
+  host_compiler_includes = _host_compiler_includes(repository_ctx, cc_fullpath)
   cuda_defines = {
            "%{cuda_include_path}": _cuda_include_path(repository_ctx,
                                                       cuda_config),
            "%{host_compiler_includes}": host_compiler_includes,
        }
-  if _use_cuda_clang(repository_ctx):
+  if is_cuda_clang:
     cuda_defines["%{clang_path}"] = cc
     _tpl(repository_ctx, "crosstool:BUILD", {"%{linker_files}": ":empty"})
     _tpl(repository_ctx, "crosstool:CROSSTOOL_clang", cuda_defines, out="crosstool/CROSSTOOL")
@@ -1046,7 +1085,10 @@
     implementation = _cuda_autoconf_impl,
     environ = [
         _GCC_HOST_COMPILER_PATH,
+        _CLANG_CUDA_COMPILER_PATH,
         "TF_NEED_CUDA",
+        "TF_CUDA_CLANG",
+        _TF_DOWNLOAD_CLANG,
         _CUDA_TOOLKIT_PATH,
         _CUDNN_INSTALL_PATH,
         _TF_CUDA_VERSION,
diff --git a/third_party/gpus/download_clang.bzl b/third_party/gpus/download_clang.bzl
new file mode 100644
index 0000000..8d235f3
--- /dev/null
+++ b/third_party/gpus/download_clang.bzl
@@ -0,0 +1,54 @@
+""" Helpers to download a recent clang release."""
+
+def _get_platform_folder(os_name):
+  os_name = os_name.lower()
+  if os_name.startswith('windows'):
+    return 'Win'
+  if os_name.startswith('mac os'):
+    return 'Mac'
+  if not os_name.startswith('linux'):
+    fail('Unknown platform')
+  return 'Linux_x64'
+
+def _download_chromium_clang(repo_ctx, platform_folder, package_version, sha256,
+                             out_folder):
+  cds_url = 'https://commondatastorage.googleapis.com/chromium-browser-clang'
+  cds_file = 'clang-%s.tgz' % package_version
+  cds_full_url = '{0}/{1}/{2}'.format(cds_url, platform_folder, cds_file)
+  repo_ctx.download_and_extract(cds_full_url, output=out_folder, sha256=sha256)
+
+def download_clang(repo_ctx, out_folder):
+  """ Download a fresh clang release and put it into out_folder.
+
+  Clang itself will be located in 'out_folder/bin/clang'.
+  We currently download one of the latest releases of clang by the
+  Chromium project (see
+  https://chromium.googlesource.com/chromium/src/+/master/docs/clang.md).
+
+  Args:
+    repo_ctx: An instance of repository_context object.
+    out_folder: A folder to extract the compiler into.
+  """
+  # TODO(ibiryukov): we currently download and extract some extra tools in the
+  # clang release (e.g., sanitizers). We should probably remove the ones
+  # we don't need and document the ones we want provide in addition to clang.
+
+  # Latest CLANG_REVISION and CLANG_SUB_REVISION of the Chromiums's release
+  # can be found in https://chromium.googlesource.com/chromium/src/tools/clang/+/master/scripts/update.py
+  CLANG_REVISION = '318667'
+  CLANG_SUB_REVISION = 1
+
+  package_version = '%s-%s' % (CLANG_REVISION, CLANG_SUB_REVISION)
+
+  checksums = {
+      'Linux_x64':
+          'e63e5fe3ec8eee4779812cd16aae0ddaf1256d2e8e93cdd5914a3d3e01355dc1',
+      'Mac':
+          '4f0aca6ec66281be94c3045550ae15a73befa59c32396112abda0030ef22e9b6',
+      'Win':
+          '7e848f2a586ea01effc51f5776dee6ffbeae0865eec6ca8a73ac9565ed299f8e',
+  }
+
+  platform_folder = _get_platform_folder(repo_ctx.os.name)
+  _download_chromium_clang(repo_ctx, platform_folder, package_version,
+                           checksums[platform_folder], out_folder)