diff --git a/paddle/fluid/operators/ngraph/ops/activation_op.h b/paddle/fluid/operators/ngraph/ops/activation_op.h
index ef6c11bce706a84c8e6a330f3acaf69f63617516..884ec659267a5386b5715b9f8b38be8900123823 100644
--- a/paddle/fluid/operators/ngraph/ops/activation_op.h
+++ b/paddle/fluid/operators/ngraph/ops/activation_op.h
@@ -26,6 +26,52 @@ namespace paddle {
 namespace operators {
 namespace ngraphs {
 
+void BuildGeluNode(
+    const std::shared_ptr<framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto input = platform::GetInputNode(op, "X", ngb_node_map);
+  auto half = paddle::platform::CreateConstant(input->get_element_type(),
+                                               input->get_shape(), {0.5});
+  auto one = paddle::platform::CreateConstant(input->get_element_type(),
+                                              input->get_shape(), {1});
+  auto sqrt_two =
+      std::make_shared<ngraph::op::Sqrt>(paddle::platform::CreateConstant(
+          input->get_element_type(), input->get_shape(), {2}));
+  auto out = half * input *
+             (one + std::make_shared<ngraph::op::Erf>(input / sqrt_two));
+  platform::SetOutputNode(op, "Out", out, ngb_node_map);
+}
+
+void BuildGeluGradNode(
+    const std::shared_ptr<framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto input = platform::GetInputNode(op, "X", ngb_node_map);
+  auto dout = platform::GetInputNode(op, "Out@GRAD", ngb_node_map);
+  auto half = paddle::platform::CreateConstant(input->get_element_type(),
+                                               input->get_shape(), {0.5});
+  auto minus_half = paddle::platform::CreateConstant(
+      input->get_element_type(), input->get_shape(), {-0.5});
+  auto one = paddle::platform::CreateConstant(input->get_element_type(),
+                                              input->get_shape(), {1});
+  auto two = paddle::platform::CreateConstant(input->get_element_type(),
+                                              input->get_shape(), {2});
+  auto pi = paddle::platform::CreateConstant(
+      input->get_element_type(), input->get_shape(), {3.14159265359});
+  auto sqrt_two = std::make_shared<ngraph::op::Sqrt>(two);
+  auto sqrt_pi = std::make_shared<ngraph::op::Sqrt>(pi);
+
+  auto first =
+      half * (one + std::make_shared<ngraph::op::Erf>(input * one / sqrt_two));
+  auto second = half * (two / sqrt_pi) * (one / sqrt_two) * input *
+                std::make_shared<ngraph::op::Exp>(minus_half * input * input);
+  auto gelu_grad = dout * (first + second);
+  platform::SetOutputNode(op, "X@GRAD", gelu_grad, ngb_node_map);
+}
+
 void BuildReluGradNode(
     const std::shared_ptr<framework::OperatorBase>& op,
     std::shared_ptr<
@@ -64,6 +110,8 @@ void BuildTanhGradNode(
 }  // namespace operators
 }  // namespace paddle
 
+REGISTER_NG_OP(gelu, BuildGeluNode);
+REGISTER_NG_OP(gelu_grad, BuildGeluGradNode);
 REGISTER_NG_OP(relu_grad, BuildReluGradNode);
 REGISTER_NG_OP(square, BuildSquareNode);
 REGISTER_NG_OP(tanh_grad, BuildTanhGradNode);
diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py
index 3c1db3bf6406cce7952b7e07dc5b39c1389d7cea..a7f167cbd415c9000311aa45bef0432f61e668ea 100644
--- a/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py
+++ b/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py
@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 import paddle.fluid.core as core
 from paddle.fluid.tests.unittests.op_test import OpTest
-from paddle.fluid.tests.unittests.test_activation_op import TestAbs, TestSigmoid, TestSquare, TestRelu, TestTanh
+from paddle.fluid.tests.unittests.test_activation_op import TestAbs, TestGelu, TestSigmoid, TestSquare, TestRelu, TestTanh
 
 
 class TestNGRAPHReluDim4(TestRelu):