diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index ef2c55cc3799ba2fac54f3c9370505b63ef22ad3..7756a52ca9b794d70c7d98ffddcec264861f3e0e 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#include <gflags/gflags.h>
 #include <glog/logging.h>
 
 #include <algorithm>
@@ -21,6 +22,10 @@ limitations under the License. */
 #include "paddle/framework/shape_inference.h"
 #include "paddle/framework/var_type.h"
 
+DEFINE_bool(op_sync, false,
+            "Default cuda is asynchronous device, set to True will"
+            "force op run in synchronous mode.");
+
 namespace paddle {
 namespace framework {
 
@@ -542,8 +547,14 @@ void OperatorWithKernel::Run(const Scope& scope,
 
   auto kernel_iter = kernels.find(expected_kernel_key);
 
-  kernel_iter->second->Compute(ExecutionContext(
-      *this, new_scope, *pool.Get(expected_kernel_key.place_)));
+  auto* new_dev_ctx = pool.Get(expected_kernel_key.place_);
+  kernel_iter->second->Compute(
+      ExecutionContext(*this, new_scope, *new_dev_ctx));
+
+  /*For profiling/benchmark only*/
+  if (FLAGS_op_sync) {
+    new_dev_ctx->Wait();
+  }
 }
 
 proto::DataType OperatorWithKernel::IndicateDataType(
diff --git a/python/paddle/v2/fluid/__init__.py b/python/paddle/v2/fluid/__init__.py
index ccd5998e3592a1f5dc795ee24875c1aed230587e..422aa0a5ba2e490d9d0667c9eff7c46f1b751b4c 100644
--- a/python/paddle/v2/fluid/__init__.py
+++ b/python/paddle/v2/fluid/__init__.py
@@ -58,7 +58,7 @@ def __bootstrap__():
 
     read_env_flags = ['use_pinned_memory', 'check_nan_inf']
     if core.is_compile_gpu():
-        read_env_flags.append('fraction_of_gpu_memory_to_use')
+        read_env_flags += ['fraction_of_gpu_memory_to_use', 'op_sync']
     core.init_gflags([sys.argv[0]] +
                      ["--tryfromenv=" + ",".join(read_env_flags)])
     core.init_glog(sys.argv[0])