diff --git a/paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc b/paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
index 0ecb6266e4a16cc5d2c22a47c0513da4410a6215..13acd3fa636804ad6534d407143cce07bdab2930 100644
--- a/paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
+++ b/paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
@@ -90,4 +90,5 @@ REGISTER_OP_KERNEL(
     ops::EltwiseMKLDNNKernel<uint8_t, dnnl::algorithm::binary_add>)
 
 REGISTER_OP_KERNEL(elementwise_add_grad, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::EltwiseAddMKLDNNGradKernel<paddle::platform::bfloat16>,
                    ops::EltwiseAddMKLDNNGradKernel<float>)
diff --git a/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_bf16_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_bf16_mkldnn_op.py
index 7e4a1172380265d2c1027b9582170ebe149d9685..ac235e00755e93efd38e4fe8b9afa4a437880ae4 100644
--- a/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_bf16_mkldnn_op.py
+++ b/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_bf16_mkldnn_op.py
@@ -30,10 +30,10 @@ class TestElementwiseAddBf16MklDNNOp(OpTest):
         self.axis = -1
 
         self.generate_data()
-        self.inputs = {
-            'X': convert_float_to_uint16(self.x),
-            'Y': convert_float_to_uint16(self.y)
-        }
+        self.x_bf16 = convert_float_to_uint16(self.x)
+        self.y_bf16 = convert_float_to_uint16(self.y)
+
+        self.inputs = {'X': self.x_bf16, 'Y': self.y_bf16}
         self.attrs = {'axis': self.axis, 'use_mkldnn': self.use_mkldnn}
         self.outputs = {'Out': convert_float_to_uint16(self.out)}
 
@@ -45,14 +45,30 @@ class TestElementwiseAddBf16MklDNNOp(OpTest):
     def test_check_output(self):
         self.check_output_with_place(core.CPUPlace())
 
+    # elementwise_add grad is just passing upper gradients to either X or Y or both
     def test_check_grad_normal(self):
-        pass
+        self.check_grad_with_place(
+            core.CPUPlace(), ["X", "Y"],
+            "Out",
+            check_dygraph=False,
+            user_defined_grads=[self.x_bf16, self.x_bf16],
+            user_defined_grad_outputs=[self.x_bf16])
 
     def test_check_grad_ingore_x(self):
-        pass
+        self.check_grad_with_place(
+            core.CPUPlace(), ["Y"],
+            "Out",
+            check_dygraph=False,
+            user_defined_grads=[self.y_bf16],
+            user_defined_grad_outputs=[self.y_bf16])
 
     def test_check_grad_ingore_y(self):
-        pass
+        self.check_grad_with_place(
+            core.CPUPlace(), ["X"],
+            "Out",
+            check_dygraph=False,
+            user_defined_grads=[self.x_bf16],
+            user_defined_grad_outputs=[self.x_bf16])
 
 
 if __name__ == '__main__':