diff --git a/paddle/fluid/operators/gelu_op_npu.cc b/paddle/fluid/operators/gelu_op_npu.cc
index 6e60926cc7951aa777138bb8785083eb48ee50dd..4db82e96cfae7c3a0332f5601b3477780c3d16d1 100644
--- a/paddle/fluid/operators/gelu_op_npu.cc
+++ b/paddle/fluid/operators/gelu_op_npu.cc
@@ -61,13 +61,14 @@ class GeluGradNPUKernel : public framework::OpKernel<T> {
         ctx.template device_context<paddle::platform::NPUDeviceContext>()
             .stream();
 
-    Tensor out(x->type());
-    out.mutable_data<T>(x->dims(), place);
-    const auto& runner_out = NpuOpRunner("Gelu", {*x}, {out}, {});
-    runner_out.Run(stream);
-
+    // NOTE(pangyoki): In the original implementation of GeluGrad op, the input
+    // is {*dout, *x, out}, where out = Gelu(x). However, we find that variable
+    // `out` was not actually used. In order to improve performance, the
+    // useless GELU operation was deleted.
+    // We directly use `*dout` as a placeholder to replace `out`, it will not
+    // be used in calculations.
     const auto& runner_dx =
-        NpuOpRunner("GeluGrad", {*dout, *x, out}, {*dx}, {});
+        NpuOpRunner("GeluGrad", {*dout, *x, *dout}, {*dx}, {});
     runner_dx.Run(stream);
   }
 };
diff --git a/python/paddle/fluid/tests/unittests/npu/test_gelu_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_gelu_op_npu.py
index efa1918206b035005dd12939b7001933266c107c..d811aaf228ddf5e264c56b23f69fc318a9352c73 100644
--- a/python/paddle/fluid/tests/unittests/npu/test_gelu_op_npu.py
+++ b/python/paddle/fluid/tests/unittests/npu/test_gelu_op_npu.py
@@ -58,12 +58,9 @@ class TestGelu(OpTest):
     def test_check_output(self):
         self.check_output_with_place(self.place, check_dygraph=False, atol=1e-3)
 
-    # TODO(ascendrc): Add grad test
-    # def test_check_grad(self):
-    #     if self.dtype == np.float16:
-    #         return
-    #     self.check_grad(['X'], 'Out')
-    #
+    def test_check_grad(self):
+        self.check_grad_with_place(
+            self.place, ['X'], 'Out', check_dygraph=False)
 
 
 @unittest.skipIf(not paddle.is_compiled_with_npu(),
@@ -115,10 +112,10 @@ class TestGeluNet(unittest.TestCase):
                 name="label", shape=[32, 1], dtype='int64')
 
             c = paddle.multiply(a, b)
-            d = fluid.layers.gelu(c)
 
-            fc_1 = fluid.layers.fc(input=d, size=128)
-            prediction = fluid.layers.fc(input=fc_1, size=2, act='softmax')
+            fc_1 = fluid.layers.fc(input=c, size=128)
+            fc_1_gelu = fluid.layers.gelu(fc_1)
+            prediction = fluid.layers.fc(input=fc_1_gelu, size=2, act='softmax')
 
             cost = fluid.layers.cross_entropy(input=prediction, label=label)
             loss = fluid.layers.reduce_mean(cost)