diff --git a/paddle/fluid/operators/fake_quantize_op.cc b/paddle/fluid/operators/fake_quantize_op.cc
index 3bb07d383548e6f4be810c96d2a916c0fe5e45f5..c873ee6718018604ae9bba74aa616b2165ba9bc4 100644
--- a/paddle/fluid/operators/fake_quantize_op.cc
+++ b/paddle/fluid/operators/fake_quantize_op.cc
@@ -134,6 +134,61 @@ $$Out = round(X/scale * range)$$
   }
 };
 
+class FakeChannelWiseQuantizeAbsMaxOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of FakeChannelWiseQuantizeOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("Out"),
+        "Output(Out) of FakeChannelWiseQuantizeOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("OutScales"),
+        "Output(Scales) of FakeChannelWiseQuantizeOp should not be null.");
+    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+    ctx->SetOutputDim("OutScales", {ctx->GetInputDim("X")[0]});
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(ctx.Input<framework::LoDTensor>("X")->type(),
+                                   ctx.GetPlace());
+  }
+};
+
+class FakeChannelWiseQuantizeAbsMaxOpMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "(Tensor) Input is float data type.");
+    AddOutput("Out",
+              "(Tensor) Output of quantized low level tensor, "
+              "but also saved as float data type.");
+    AddOutput("OutScales", "(Tensor) Current channel wise scale");
+    AddAttr<int>("bit_length", "(int, default 8)")
+        .SetDefault(8)
+        .AddCustomChecker([](const int& bit_length) {
+          PADDLE_ENFORCE(bit_length >= 1 && bit_length <= 16,
+                         "'bit_length' should be between 1 and 16.");
+        });
+    AddComment(R"DOC(
+The scale of FakeChannelWiseQuantize operator is a vector.
+In detail, each channel of the input X has a scale value.
+
+$$scale_c = max(abs(X_c))$$
+$$range = 2^{bit_length - 1} - 1$$
+$$Out_c = round(X_c / scale_c * range)$$
+
+In above three formulas, the range value of c is as follow:
+$$0 \leq c \leq \ the\ channel\ number\ of\ X$$
+)DOC");
+  }
+};
+
 class FakeQuantizeRangeAbsMaxOp : public framework::OperatorWithKernel {
  public:
   FakeQuantizeRangeAbsMaxOp(const std::string& type,
@@ -218,3 +273,10 @@ REGISTER_OPERATOR(fake_quantize_range_abs_max, ops::FakeQuantizeRangeAbsMaxOp,
                   paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(fake_quantize_range_abs_max,
                        ops::FakeQuantizeRangeAbsMaxKernel<CPU, float>);
+
+REGISTER_OPERATOR(fake_channel_wise_quantize_abs_max,
+                  ops::FakeChannelWiseQuantizeAbsMaxOp,
+                  ops::FakeChannelWiseQuantizeAbsMaxOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(fake_channel_wise_quantize_abs_max,
+                       ops::FakeChannelWiseQuantizeAbsMaxKernel<CPU, float>);
diff --git a/paddle/fluid/operators/fake_quantize_op.cu b/paddle/fluid/operators/fake_quantize_op.cu
index a0ff6396210c2b3a7f8bd6b9f274b875d7fd4933..5da16a7c7314c62034bff67bcc8d099e2799c3de 100644
--- a/paddle/fluid/operators/fake_quantize_op.cu
+++ b/paddle/fluid/operators/fake_quantize_op.cu
@@ -174,5 +174,7 @@ namespace ops = paddle::operators;
 using CUDA = paddle::platform::CUDADeviceContext;
 REGISTER_OP_CUDA_KERNEL(fake_quantize_abs_max,
                         ops::FakeQuantizeAbsMaxKernel<CUDA, float>);
+REGISTER_OP_CUDA_KERNEL(fake_channel_wise_quantize_abs_max,
+                        ops::FakeChannelWiseQuantizeAbsMaxKernel<CUDA, float>);
 REGISTER_OP_CUDA_KERNEL(fake_quantize_range_abs_max,
                         ops::FakeQuantizeRangeAbsMaxKernel<CUDA, float>);
diff --git a/paddle/fluid/operators/fake_quantize_op.h b/paddle/fluid/operators/fake_quantize_op.h
index 7ace7573ec5c03ab8788cfc0aab614b7f80ea073..8b47600e7d99ad9e4e40ae162582d4c8461224ad 100644
--- a/paddle/fluid/operators/fake_quantize_op.h
+++ b/paddle/fluid/operators/fake_quantize_op.h
@@ -63,6 +63,39 @@ class FakeQuantizeAbsMaxKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class FakeChannelWiseQuantizeAbsMaxKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in = context.Input<framework::Tensor>("X");
+
+    auto* out = context.Output<framework::Tensor>("Out");
+    auto* out_scales = context.Output<framework::Tensor>("OutScales");
+    T* out_scales_data = out_scales->mutable_data<T>(context.GetPlace());
+    out->mutable_data<T>(context.GetPlace());
+
+    int bit_length = context.Attr<int>("bit_length");
+    int bin_cnt = std::pow(2, bit_length - 1) - 1;
+
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    auto find_abs_max = FindAbsMaxFunctor<DeviceContext, T>();
+    for (int64_t i = 0; i < in->dims()[0]; i++) {
+      framework::Tensor one_channel = in->Slice(i, i + 1);
+      const T* one_channel_data = one_channel.data<T>();
+      find_abs_max(dev_ctx, one_channel_data, one_channel.numel(),
+                   &out_scales_data[i]);
+    }
+    auto clip_quant = ClipAndFakeQuantFunctor<DeviceContext, T>();
+    for (int64_t i = 0; i < in->dims()[0]; i++) {
+      framework::Tensor one_channel_in = in->Slice(i, i + 1);
+      framework::Tensor one_channel_out = out->Slice(i, i + 1);
+      framework::Tensor one_channel_scale = out_scales->Slice(i, i + 1);
+      clip_quant(dev_ctx, one_channel_in, one_channel_scale, bin_cnt,
+                 &one_channel_out);
+    }
+  }
+};
+
 template <typename DeviceContext, typename T>
 class FakeQuantizeRangeAbsMaxKernel : public framework::OpKernel<T> {
  public:
diff --git a/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py b/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py
index 4582b2a0eed401235835374d4cd58782d8d3a68f..90a90112bd5f0e24374111073514b20dd1231edb 100644
--- a/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py
+++ b/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py
@@ -35,6 +35,30 @@ class TestFakeQuantizeOp(OpTest):
         self.check_output()
 
 
+class TestFakeChannelWiseQuantizeOp(OpTest):
+    def setUp(self):
+        self.op_type = "fake_channel_wise_quantize_abs_max"
+        self.attrs = {'bit_length': 8}
+        self.inputs = {
+            'X': np.random.random((4, 3, 64, 64)).astype("float32"),
+        }
+        scales = []
+        for i in range(self.inputs['X'].shape[0]):
+            scales.append(np.max(np.abs(self.inputs['X'][i])).astype("float32"))
+        outputs = self.inputs['X'].copy()
+        for i, scale in enumerate(scales):
+            outputs[i] = np.round(outputs[i] / scale * (
+                (1 << (self.attrs['bit_length'] - 1)) - 1))
+
+        self.outputs = {
+            'Out': outputs,
+            'OutScales': np.array(scales).astype("float32"),
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
 class TestFakeQuantizeRangeAbsMaxOp(OpTest):
     def setUp(self):
         self.op_type = "fake_quantize_range_abs_max"