diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 5a100c5746e616e860811dd47da27036ea7355d5..6a28b975f8e3fc7c7bd56f3a88464cbe4b4c9a72 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -45,7 +45,7 @@ ENDIF()
 cc_library(cpu_info SRCS cpu_info.cc DEPS ${CPU_INFO_DEPS})
 cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info)
 
-nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce monitor)
+nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce monitor dynload_cuda)
 
 cc_library(place SRCS place.cc DEPS enforce boost)
 cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
diff --git a/paddle/fluid/platform/gpu_info.cc b/paddle/fluid/platform/gpu_info.cc
index 5f63233d8bee4beefd6e1695d8bc3d6e5e4ae7fb..ca1e5501c6a84e6136c28f564a78a7e63f0ee8d4 100644
--- a/paddle/fluid/platform/gpu_info.cc
+++ b/paddle/fluid/platform/gpu_info.cc
@@ -19,6 +19,7 @@ limitations under the License. */
 
 #include "gflags/gflags.h"
 #include "paddle/fluid/platform/cuda_device_guard.h"
+#include "paddle/fluid/platform/dynload/cudnn.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/lock_guard_ptr.h"
 #include "paddle/fluid/platform/macros.h"
@@ -38,11 +39,11 @@ USE_GPU_MEM_STAT;
 namespace paddle {
 namespace platform {
 
-/* Here is a very simple CUDA “pro tip”: cudaDeviceGetAttribute() is a much
-faster way to query device properties. You can see details in
-https://devblogs.nvidia.com/cuda-pro-tip-the-fast-way-to-query-device-properties/
-*/
+int CudnnVersion() {
+  if (!dynload::HasCUDNN()) return -1;
 
+  return dynload::cudnnGetVersion();
+}
 static int GetCUDADeviceCountImpl() {
   int driverVersion = 0;
   cudaError_t status = cudaDriverGetVersion(&driverVersion);
@@ -73,6 +74,10 @@ int GetCUDADeviceCount() {
   return dev_cnt;
 }
 
+/* Here is a very simple CUDA “pro tip”: cudaDeviceGetAttribute() is a much
+faster way to query device properties. You can see details in
+https://devblogs.nvidia.com/cuda-pro-tip-the-fast-way-to-query-device-properties/
+*/
 int GetCUDAComputeCapability(int id) {
   PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(),
                     platform::errors::InvalidArgument(
diff --git a/paddle/fluid/platform/gpu_info.h b/paddle/fluid/platform/gpu_info.h
index 6a9893647172e2c63f4749fdb0ae1cb0fdfaaf04..ec77447ef77dbb1cd7ee180176f95a9ab8f7c03a 100644
--- a/paddle/fluid/platform/gpu_info.h
+++ b/paddle/fluid/platform/gpu_info.h
@@ -23,6 +23,8 @@ limitations under the License. */
 
 namespace paddle {
 namespace platform {
+//! Get the version of cudnn
+int CudnnVersion();
 
 //! Get the total number of GPU devices in system.
 int GetCUDADeviceCount();
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 7667bb50aa49fc9fbe32612695fd93e8676cacf4..f426ca82966d3ec39e5504660c394c4539a39d5e 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -341,6 +341,10 @@ PYBIND11_MODULE(core_noavx, m) {
 
   m.def("set_num_threads", &platform::SetNumThreads);
 
+#ifdef PADDLE_WITH_CUDA
+  m.def("cudnn_version", &platform::CudnnVersion);
+#endif
+
   m.def("from_dlpack", [](py::capsule *dltensor) {
     DLManagedTensor *dmt = reinterpret_cast<DLManagedTensor *>(
         PyCapsule_GetPointer(dltensor->ptr(), "dltensor"));
diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py
index d62d8c789a71ee4ccc12542164c4ebe416500d81..6235f1c12ba7e9b624d905dcc0de46a794afe6d2 100644
--- a/python/paddle/__init__.py
+++ b/python/paddle/__init__.py
@@ -232,6 +232,7 @@ from .tensor.stat import reduce_mean  #DEFINE_ALIAS
 from .tensor.stat import std  #DEFINE_ALIAS
 from .tensor.stat import var  #DEFINE_ALIAS
 from .fluid.data import data
+from .device import get_cudnn_version
 from .device import set_device
 from .device import get_device
 # from .tensor.tensor import Tensor        #DEFINE_ALIAS
diff --git a/python/paddle/device.py b/python/paddle/device.py
index 25b4ebe55609ddd36a22678fc2689886b231050f..e2ef8e7092ad3f6af91c8d5d3c0b1deaed025514 100644
--- a/python/paddle/device.py
+++ b/python/paddle/device.py
@@ -12,10 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# TODO: define the functions to manipulate devices 
 from paddle.fluid import core
 from paddle.fluid import framework
 import re
+
 __all__ = [
+    'get_cudnn_version',
     'set_device',
     'get_device'
     #            'cpu_places',
@@ -27,6 +30,40 @@ __all__ = [
     #            'is_compiled_with_cuda'
 ]
 
+_cudnn_version = None
+
+
+def get_cudnn_version():
+    """
+    This funciton return the version of cudnn. the retuen value is int which represents the 
+    cudnn version. For example, if it return 7600, it represents the version of cudnn is 7.6.
+    
+    Returns:
+        int: A int value which represents the cudnn version. If cudnn version is not installed, it return None.
+
+    Examples:
+        .. code-block:: python
+            
+            import paddle
+
+            cudnn_version = get_cudnn_version()
+
+
+
+    """
+    global _cudnn_version
+    if not core.is_compiled_with_cuda():
+        return None
+    if _cudnn_version is None:
+        cudnn_version = int(core.cudnn_version())
+        _cudnn_version = cudnn_version
+        if _cudnn_version < 0:
+            return None
+        else:
+            return cudnn_version
+    else:
+        return _cudnn_version
+
 
 def set_device(device):
     """
diff --git a/python/paddle/fluid/tests/unittests/test_query_op.py b/python/paddle/fluid/tests/unittests/test_query_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..fc8ce5ad5f6b89b28fb2ddddd15d5b315fe4c0e4
--- /dev/null
+++ b/python/paddle/fluid/tests/unittests/test_query_op.py
@@ -0,0 +1,32 @@
+#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# 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.
+
+from __future__ import print_function
+
+import unittest
+import paddle
+from paddle.fluid import core
+
+
+class TestCudnnVersion(unittest.TestCase):
+    def test_no_cudnn(self):
+        cudnn_version = paddle.get_cudnn_version()
+        if not core.is_compiled_with_cuda():
+            self.assertEqual((cudnn_version is None), True)
+        else:
+            self.assertEqual((isinstance(cudnn_version, int)), True)
+
+
+if __name__ == '__main__':
+    unittest.main()