diff --git a/paddle/framework/init.cc b/paddle/framework/init.cc
index 4deb4fa903dec04e9b76c5a620f1eb76c9f1db07..3ff2da344627ed3ada3955ec5ee2c886402554f4 100644
--- a/paddle/framework/init.cc
+++ b/paddle/framework/init.cc
@@ -54,7 +54,7 @@ bool InitDevices(const std::vector<std::string> &devices) {
 #ifdef PADDLE_WITH_CUDA
       auto pos = string::RFind(p, ':', string::Piece::npos);
       auto number = device.substr(pos + 1);
-      places.emplace_back(platform::GPUPlace(std::stoi(number)));
+      places.emplace_back(platform::CUDAPlace(std::stoi(number)));
 #else
       LOG(WARNING)
           << "'GPU' is not supported, Please re-compile with WITH_GPU option";
diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc
index 465f8c62b5fe2efd549f68bb3a9823d299ba5393..d766d3c4163b6b7c6fdc772acb4b7e7b315f8783 100644
--- a/paddle/framework/lod_tensor.cc
+++ b/paddle/framework/lod_tensor.cc
@@ -224,7 +224,7 @@ void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
       while (size != 0) {
         size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
         memory::Copy(cpu, buf.get(),
-                     boost::get<platform::GPUPlace>(tensor.place()),
+                     boost::get<platform::CUDAPlace>(tensor.place()),
                      reinterpret_cast<const void *>(data), size_to_write,
                      gpu_dev_ctx.stream());
         gpu_dev_ctx.Wait();
diff --git a/paddle/framework/lod_tensor_test.cu b/paddle/framework/lod_tensor_test.cu
index 5b90fbfca7f6bec4f2c862d0ff18dfd7cf39e181..e8508ad2658ae850e4c98aa798b5db6d007e67d0 100644
--- a/paddle/framework/lod_tensor_test.cu
+++ b/paddle/framework/lod_tensor_test.cu
@@ -27,7 +27,7 @@ __global__ void test(size_t* a, int size) {
 
 TEST(LoDTensor, LoDInGPU) {
   paddle::framework::LoDTensor lod_tensor;
-  paddle::platform::GPUPlace place(0);
+  paddle::platform::CUDAPlace place(0);
 
   paddle::framework::LoD src_lod;
   src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
diff --git a/paddle/framework/op_kernel_type_test.cc b/paddle/framework/op_kernel_type_test.cc
index 899676b5c1a3799427d6ef03bae47284550c781b..8753d7cc378662ce116e447dc6a340a07e5dd2ca 100644
--- a/paddle/framework/op_kernel_type_test.cc
+++ b/paddle/framework/op_kernel_type_test.cc
@@ -37,13 +37,13 @@ TEST(OpKernelType, Hash) {
   using OpKernelType = paddle::framework::OpKernelType;
   using DataType = paddle::framework::proto::DataType;
   using CPUPlace = paddle::platform::CPUPlace;
-  using GPUPlace = paddle::platform::GPUPlace;
+  using CUDAPlace = paddle::platform::CUDAPlace;
   using DataLayout = paddle::framework::DataLayout;
   using LibraryType = paddle::framework::LibraryType;
 
   OpKernelType op_kernel_type_1(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
                                 LibraryType::kCUDNN);
-  OpKernelType op_kernel_type_2(DataType::FP32, GPUPlace(0), DataLayout::kNCHW,
+  OpKernelType op_kernel_type_2(DataType::FP32, CUDAPlace(0), DataLayout::kNCHW,
                                 LibraryType::kCUDNN);
 
   OpKernelType::Hash hasher;
diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index 244c1174655f61cad7176a211b07863dbfbba9aa..9bb2a3b5c2931d03152cc3262c0ad8da17b8aacb 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -188,7 +188,7 @@ class OpKernelRegistrar : public Registrar {
   }
 
 #define REGISTER_OP_CUDA_KERNEL(op_type, ...) \
-  REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::GPUPlace, __VA_ARGS__)
+  REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__)
 
 #define REGISTER_OP_CPU_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
diff --git a/paddle/framework/tensor.md b/paddle/framework/tensor.md
index 7a80816d8e4ffa3a9462f3d9b87eff0f048466aa..0a27ac9bb6b03649d42e12100fda9e80a56e7f56 100644
--- a/paddle/framework/tensor.md
+++ b/paddle/framework/tensor.md
@@ -71,7 +71,7 @@ private:
 ```
 
 ```c++
-typedef boost::variant<GpuPlace, CpuPlace> Place;
+typedef boost::variant<CUDAPlace, CpuPlace> Place;
 typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>,
                        Dim<6>, Dim<7>, Dim<8>, Dim<9>> DDimVar;
 typedef boost::variant<
diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h
index aba1f9f09329f890ef190f8820b958c56f017e89..3d93b7808bc96143b5261e1ec41ad98b15c74975 100644
--- a/paddle/framework/tensor_impl.h
+++ b/paddle/framework/tensor_impl.h
@@ -125,11 +125,11 @@ inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
           boost::get<platform::CPUPlace>(place), size, type));
     } else if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
-      PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
+      PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
     }
 #else
-      holder_.reset(new PlaceholderImpl<platform::GPUPlace>(
-          boost::get<platform::GPUPlace>(place), size, type));
+      holder_.reset(new PlaceholderImpl<platform::CUDAPlace>(
+          boost::get<platform::CUDAPlace>(place), size, type));
     }
 #endif
     offset_ = 0;
diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc
index ceca64365a1a628642eb374a3e3bbdff490c955a..f347981f2e11fdc4770df8d5e7277cd55744ab3f 100644
--- a/paddle/framework/tensor_test.cc
+++ b/paddle/framework/tensor_test.cc
@@ -80,20 +80,20 @@ TEST(Tensor, MutableData) {
     float* p1 = nullptr;
     float* p2 = nullptr;
     // initialization
-    p1 = src_tensor.mutable_data<float>(make_ddim({1, 2, 3}), GPUPlace());
+    p1 = src_tensor.mutable_data<float>(make_ddim({1, 2, 3}), CUDAPlace());
     EXPECT_NE(p1, nullptr);
     // set src_tensor a new dim with large size
     // momery is supposed to be re-allocated
-    p2 = src_tensor.mutable_data<float>(make_ddim({3, 4}), GPUPlace());
+    p2 = src_tensor.mutable_data<float>(make_ddim({3, 4}), CUDAPlace());
     EXPECT_NE(p2, nullptr);
     EXPECT_NE(p1, p2);
     // set src_tensor a new dim with same size
     // momery block is supposed to be unchanged
-    p1 = src_tensor.mutable_data<float>(make_ddim({2, 2, 3}), GPUPlace());
+    p1 = src_tensor.mutable_data<float>(make_ddim({2, 2, 3}), CUDAPlace());
     EXPECT_EQ(p1, p2);
     // set src_tensor a new dim with smaller size
     // momery block is supposed to be unchanged
-    p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), GPUPlace());
+    p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), CUDAPlace());
     EXPECT_EQ(p1, p2);
   }
 #endif
@@ -130,7 +130,7 @@ TEST(Tensor, ShareDataWith) {
   {
     Tensor src_tensor;
     Tensor dst_tensor;
-    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), GPUPlace());
+    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), CUDAPlace());
     dst_tensor.ShareDataWith(src_tensor);
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
@@ -166,7 +166,7 @@ TEST(Tensor, Slice) {
 #ifdef PADDLE_WITH_CUDA
   {
     Tensor src_tensor;
-    src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
+    src_tensor.mutable_data<double>(make_ddim({6, 9}), CUDAPlace());
     Tensor slice_tensor = src_tensor.Slice(2, 6);
     DDim slice_dims = slice_tensor.dims();
     ASSERT_EQ(arity(slice_dims), 2);
@@ -176,11 +176,11 @@ TEST(Tensor, Slice) {
     uintptr_t src_data_address =
         reinterpret_cast<uintptr_t>(src_tensor.data<double>());
     uintptr_t src_mutable_data_address = reinterpret_cast<uintptr_t>(
-        src_tensor.mutable_data<double>(src_tensor.dims(), GPUPlace()));
+        src_tensor.mutable_data<double>(src_tensor.dims(), CUDAPlace()));
     uintptr_t slice_data_address =
         reinterpret_cast<uintptr_t>(slice_tensor.data<double>());
     uintptr_t slice_mutable_data_address = reinterpret_cast<uintptr_t>(
-        slice_tensor.mutable_data<double>(slice_tensor.dims(), GPUPlace()));
+        slice_tensor.mutable_data<double>(slice_tensor.dims(), CUDAPlace()));
     EXPECT_EQ(src_data_address, src_mutable_data_address);
     EXPECT_EQ(slice_data_address, slice_mutable_data_address);
     EXPECT_EQ(src_data_address + 9 * 2 * sizeof(double), slice_data_address);
diff --git a/paddle/framework/tensor_util.h b/paddle/framework/tensor_util.h
index 4e34b90d57eed8fea84b83045df61a98483c8849..5b474e4aef6d8ef714a59052abd55bc439605f6d 100644
--- a/paddle/framework/tensor_util.h
+++ b/paddle/framework/tensor_util.h
@@ -47,11 +47,11 @@ inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(src_place) &&  // NOLINT
            platform::is_cpu_place(dst_place)) {
-    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
+    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_cpu_place = boost::get<platform::CPUPlace>(dst_place);
     auto ctx_place = ctx.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
-    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
     memory::Copy(
         dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size,
@@ -59,21 +59,21 @@ inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
   } else if (platform::is_cpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
     auto src_cpu_place = boost::get<platform::CPUPlace>(src_place);
-    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     auto ctx_place = ctx.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
-    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place);
     memory::Copy(
         dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
-    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
-    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
+    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     auto ctx_place = ctx.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
-    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
     PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
     memory::Copy(
         dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
@@ -108,7 +108,7 @@ inline void CopyFromVector(const std::vector<T>& src,
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(dst_place)) {  // NOLINT
     memory::Copy(
-        boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place, src_ptr,
+        boost::get<platform::CUDAPlace>(dst_place), dst_ptr, src_place, src_ptr,
         size,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   }
@@ -141,7 +141,7 @@ inline void CopyToVector(const Tensor& src, const platform::DeviceContext& ctx,
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(src.place())) {  // NOLINT
     memory::Copy(
-        dst_place, dst_ptr, boost::get<platform::GPUPlace>(src.place()),
+        dst_place, dst_ptr, boost::get<platform::CUDAPlace>(src.place()),
         src_ptr, size,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   }
diff --git a/paddle/framework/tensor_util_test.cc b/paddle/framework/tensor_util_test.cc
index 03a70de182d0eb499a81413d38229c81c4378b91..3afb98a4a52f8cff335833cfb589a87b2d8d2ef2 100644
--- a/paddle/framework/tensor_util_test.cc
+++ b/paddle/framework/tensor_util_test.cc
@@ -58,7 +58,7 @@ TEST(CopyFrom, Tensor) {
     memcpy(src_ptr, arr, 9 * sizeof(int));
 
     // CPU Tensor to GPU Tensor
-    auto gpu_place = new platform::GPUPlace(0);
+    auto gpu_place = new platform::CUDAPlace(0);
     platform::CUDADeviceContext gpu_ctx(*gpu_place);
     CopyFrom(src_tensor, *gpu_place, gpu_ctx, &gpu_tensor);
 
@@ -143,7 +143,7 @@ TEST(CopyFromVector, Tensor) {
 
     // Copy to GPUTensor
     gpu_tensor.Resize(make_ddim({3, 3}));
-    auto gpu_place = new paddle::platform::GPUPlace();
+    auto gpu_place = new paddle::platform::CUDAPlace();
     CUDADeviceContext gpu_ctx(*gpu_place);
     CopyFromVector<int>(src_vec, gpu_ctx, &gpu_tensor);
     // Copy from GPU to CPU tensor for comparison
@@ -210,7 +210,7 @@ TEST(CopyToVector, Tensor) {
   {
     std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
     Tensor gpu_tensor;
-    GPUPlace place;
+    CUDAPlace place;
     CUDADeviceContext gpu_ctx(place);
     CopyFromVector<int>(src_vec, gpu_ctx, &gpu_tensor);
 
diff --git a/paddle/memory/README.md b/paddle/memory/README.md
index 6cb003c50bc7d142d65b0591e7e5235431d2ea42..7cf61d089b39041b7a15184e0ea9211d14a66f5e 100644
--- a/paddle/memory/README.md
+++ b/paddle/memory/README.md
@@ -12,13 +12,13 @@ p = memory::Alloc(platform::CPUPlace(), 4*1024);
 To allocate 4KB memory on the 3rd GPU:
 
 ```cpp
-p = memory::Alloc(platform::GPUPlace(2), 4*1024);
+p = memory::Alloc(platform::CUDAPlace(2), 4*1024);
 ```
 
 To free memory and check the so-far used amount of memory on a place:
 
 ```cpp
-auto pl = platform::GPUPlace(0);
+auto pl = platform::CUDAPlace(0);
 p = memory::Alloc(pl, 4*1024);
 cout << memory::Used(pl);
 memory::Free(pl, p);
@@ -36,7 +36,7 @@ template <typename Place> size_t Used(Place);
 }  // namespace memory
 ```
 
-These function templates have specializations on either `platform::CPUPlace` or `platform::GPUPlace`:
+These function templates have specializations on either `platform::CPUPlace` or `platform::CUDAPlace`:
 
 ```cpp
 template<>
@@ -49,7 +49,7 @@ and
 
 ```cpp
 template<>
-void Alloc<GPUPlace>(GPUPlace p, size_t size) {
+void Alloc<CUDAPlace>(CUDAPlace p, size_t size) {
   return GetGPUBuddyAllocator(p.id)->Alloc(size);
 }
 ```
@@ -122,7 +122,7 @@ There are two implementations of `Context`:
 
 1. [`CPUContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L105), whose [`New` method](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L131) calls [`g_cpu_allocator.get()->New(size_t)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.cc#L15) to allocate the memory.
 
-1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202).  This looks very similar to class `majel::GPUPlace`, who also has an `int id_` data member.   `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory.
+1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202).  This looks very similar to class `majel::CUDAPlace`, who also has an `int id_` data member.   `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory.
 
 ### Majel
 
diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc
index 5c629dc3d2aca2705e439df836214c1284b31c8f..b46141aafd7146bd3def12d86108c10f1f143d20 100644
--- a/paddle/memory/memcpy.cc
+++ b/paddle/memory/memcpy.cc
@@ -28,31 +28,25 @@ void Copy<platform::CPUPlace, platform::CPUPlace>(platform::CPUPlace, void* dst,
 
 #ifdef PADDLE_WITH_CUDA
 template <>
-void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num,
-                                                  cudaStream_t stream) {
+void Copy<platform::CPUPlace, platform::CUDAPlace>(
+    platform::CPUPlace dst_place, void* dst, platform::CUDAPlace src_place,
+    const void* src, size_t num, cudaStream_t stream) {
   platform::SetDeviceId(src_place.device);
   platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToHost, stream);
 }
 
 template <>
-void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::CPUPlace src_place,
-                                                  const void* src, size_t num,
-                                                  cudaStream_t stream) {
+void Copy<platform::CUDAPlace, platform::CPUPlace>(
+    platform::CUDAPlace dst_place, void* dst, platform::CPUPlace src_place,
+    const void* src, size_t num, cudaStream_t stream) {
   platform::SetDeviceId(dst_place.device);
   platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyHostToDevice, stream);
 }
 
 template <>
-void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num,
-                                                  cudaStream_t stream) {
+void Copy<platform::CUDAPlace, platform::CUDAPlace>(
+    platform::CUDAPlace dst_place, void* dst, platform::CUDAPlace src_place,
+    const void* src, size_t num, cudaStream_t stream) {
   if (dst_place == src_place) {
     platform::SetDeviceId(src_place.device);
     platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToDevice, stream);
diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc
index 9cafdfda75d0511227ef648d50a8635320a81d32..c4bb6baee7ebf2941cee5915ca2723c298689261 100644
--- a/paddle/memory/memory.cc
+++ b/paddle/memory/memory.cc
@@ -83,12 +83,12 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
 }
 
 template <>
-size_t Used<platform::GPUPlace>(platform::GPUPlace place) {
+size_t Used<platform::CUDAPlace>(platform::CUDAPlace place) {
   return GetGPUBuddyAllocator(place.device)->Used();
 }
 
 template <>
-void* Alloc<platform::GPUPlace>(platform::GPUPlace place, size_t size) {
+void* Alloc<platform::CUDAPlace>(platform::CUDAPlace place, size_t size) {
   auto* buddy_allocator = GetGPUBuddyAllocator(place.device);
   auto* ptr = buddy_allocator->Alloc(size);
   if (ptr == nullptr) {
@@ -101,14 +101,14 @@ void* Alloc<platform::GPUPlace>(platform::GPUPlace place, size_t size) {
     LOG(WARNING) << "total " << total;
     LOG(WARNING) << "GpuMinChunkSize " << platform::GpuMinChunkSize();
     LOG(WARNING) << "GpuMaxChunkSize " << platform::GpuMaxChunkSize();
-    LOG(WARNING) << "GPU memory used: " << Used<platform::GPUPlace>(place);
+    LOG(WARNING) << "GPU memory used: " << Used<platform::CUDAPlace>(place);
     platform::SetDeviceId(cur_dev);
   }
   return ptr;
 }
 
 template <>
-void Free<platform::GPUPlace>(platform::GPUPlace place, void* p) {
+void Free<platform::CUDAPlace>(platform::CUDAPlace place, void* p) {
   GetGPUBuddyAllocator(place.device)->Free(p);
 }
 
diff --git a/paddle/memory/memory_test.cc b/paddle/memory/memory_test.cc
index 2444931e26774ae80b916fbb7bd46ff93025d9ed..f476bf71264da59a5c546968f4689145e1d8801b 100644
--- a/paddle/memory/memory_test.cc
+++ b/paddle/memory/memory_test.cc
@@ -82,7 +82,7 @@ TEST(BuddyAllocator, CPUMultAlloc) {
 
 #ifdef PADDLE_WITH_CUDA
 
-size_t align(size_t size, paddle::platform::GPUPlace place) {
+size_t align(size_t size, paddle::platform::CUDAPlace place) {
   size += sizeof(paddle::memory::detail::Metadata);
   size_t alignment = paddle::platform::GpuMinChunkSize();
   size_t remaining = size % alignment;
@@ -94,7 +94,7 @@ TEST(BuddyAllocator, GPUAllocation) {
 
   EXPECT_EQ(p, nullptr);
 
-  paddle::platform::GPUPlace gpu(0);
+  paddle::platform::CUDAPlace gpu(0);
   p = paddle::memory::Alloc(gpu, 4096);
 
   EXPECT_NE(p, nullptr);
@@ -103,7 +103,7 @@ TEST(BuddyAllocator, GPUAllocation) {
 }
 
 TEST(BuddyAllocator, GPUMultAlloc) {
-  paddle::platform::GPUPlace gpu;
+  paddle::platform::CUDAPlace gpu;
 
   std::unordered_map<void *, size_t> ps;
 
diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu
index dd51aad105fecf4e3118f03e2f1868abb5523bc8..0aadd5af41531e54b357756441f92da668d4ec01 100644
--- a/paddle/operators/accuracy_op.cu
+++ b/paddle/operators/accuracy_op.cu
@@ -56,7 +56,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* inference = ctx.Input<Tensor>("Out");
     auto* indices = ctx.Input<Tensor>("Indices");
     auto* label = ctx.Input<Tensor>("Label");
diff --git a/paddle/operators/batch_norm_op.cu.cc b/paddle/operators/batch_norm_op.cu.cc
index 55d0736a4c8e09eea637e5ab7e49af9a618e7fd8..3d17725ab47682355b2093782848849857f9bf59 100644
--- a/paddle/operators/batch_norm_op.cu.cc
+++ b/paddle/operators/batch_norm_op.cu.cc
@@ -53,7 +53,7 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
     const float momentum = ctx.Attr<float>("momentum");
     const bool is_test = ctx.Attr<bool>("is_test");
@@ -179,7 +179,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
     const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
     const DataLayout data_layout =
diff --git a/paddle/operators/conv_cudnn_op.cu.cc b/paddle/operators/conv_cudnn_op.cu.cc
index 3da0a9001aafbb5b2c4b9a91c4527d9437ac38a1..79e020b7556e1e349d40820616798a7a10f3b221 100644
--- a/paddle/operators/conv_cudnn_op.cu.cc
+++ b/paddle/operators/conv_cudnn_op.cu.cc
@@ -36,7 +36,7 @@ class CudnnConvOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* input = ctx.Input<Tensor>("Input");
     auto* filter = ctx.Input<Tensor>("Filter");
     auto* output = ctx.Output<Tensor>("Output");
@@ -130,7 +130,7 @@ class CudnnConvOpKernel : public framework::OpKernel<T> {
         handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
         cudnn_output_desc, algo, &workspace_size_in_bytes));
     // Allocate on GPU memory
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
     // ------------------- cudnn conv forward ---------------------
     T alpha = 1.0f, beta = 0.0f;
@@ -151,7 +151,7 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto input = ctx.Input<Tensor>("Input");
     auto filter = ctx.Input<Tensor>("Filter");
     auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
@@ -277,7 +277,7 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv workspace ---------------------
     // Already on GPU
     void* cudnn_workspace = nullptr;
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
     // ------------------- cudnn conv backward data ---------------------
     T alpha = 1.0f, beta = 0.0f;
diff --git a/paddle/operators/conv_transpose_cudnn_op.cu.cc b/paddle/operators/conv_transpose_cudnn_op.cu.cc
index f0297f6c40c132c28b50184997d657451f26362b..b3663209ff989c1d8fddd2f393db44de23e1e2a7 100644
--- a/paddle/operators/conv_transpose_cudnn_op.cu.cc
+++ b/paddle/operators/conv_transpose_cudnn_op.cu.cc
@@ -35,7 +35,7 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* input = ctx.Input<Tensor>("Input");
     auto* filter = ctx.Input<Tensor>("Filter");
     auto* output = ctx.Output<Tensor>("Output");
@@ -100,7 +100,7 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel<T> {
             cudnn_output_desc, algo, &workspace_size_in_bytes));
 
     // Allocate on GPU memory
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
 
     // ------------------- cudnn conv transpose forward ---------------------
@@ -120,7 +120,7 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto input = ctx.Input<Tensor>("Input");
     auto filter = ctx.Input<Tensor>("Filter");
     auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
@@ -201,7 +201,7 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv workspace ---------------------
     // Already on GPU
     void* cudnn_workspace = nullptr;
-    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
     // ------------------- cudnn conv backward data ---------------------
     // FIXME(typhoonzero): template type T may not be the same as cudnn call.
diff --git a/paddle/operators/detail/strided_memcpy.h b/paddle/operators/detail/strided_memcpy.h
index 068c82f399316a1587d7322d8dab75823656800e..b81bb8ba7e9f7d29a75d56b6f403bd0d7ed86f7c 100644
--- a/paddle/operators/detail/strided_memcpy.h
+++ b/paddle/operators/detail/strided_memcpy.h
@@ -35,7 +35,7 @@ struct StridedMemcpyFunctor<T, 1> {
       memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim.head);
     } else {
 #ifdef PADDLE_WITH_CUDA
-      auto& gpu_place = boost::get<platform::GPUPlace>(place);
+      auto& gpu_place = boost::get<platform::CUDAPlace>(place);
       auto& cuda_ctx =
           reinterpret_cast<const platform::CUDADeviceContext&>(dev_ctx);
       memory::Copy(gpu_place, dst, gpu_place, src, sizeof(T) * dst_dim.head,
diff --git a/paddle/operators/linear_chain_crf_op.h b/paddle/operators/linear_chain_crf_op.h
index 694584e79c3a1e818814a4a2145f52d8db7cf10a..19c6715ec877dea6dcf0babc7373333a4d9eed0f 100644
--- a/paddle/operators/linear_chain_crf_op.h
+++ b/paddle/operators/linear_chain_crf_op.h
@@ -219,8 +219,8 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> {
     // operators runs on GPU device.
     auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor& src,
                          Tensor* dst) {
-      dst->mutable_data<T>(platform::GPUPlace());
-      framework::CopyFrom(src, platform::GPUPlace(), ctx, dst);
+      dst->mutable_data<T>(platform::CUDAPlace());
+      framework::CopyFrom(src, platform::CUDAPlace(), ctx, dst);
     };
     copyTensor(ctx, emission_exps_src, emission_exps_dst);
     copyTensor(ctx, transition_exps_src, transition_exps_dst);
@@ -433,8 +433,8 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
     auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor* src,
                          Tensor* dst) {
       if (src && dst) {
-        dst->mutable_data<T>(platform::GPUPlace());
-        framework::CopyFrom(*src, platform::GPUPlace(), ctx, dst);
+        dst->mutable_data<T>(platform::CUDAPlace());
+        framework::CopyFrom(*src, platform::CUDAPlace(), ctx, dst);
       }
     };
     copyTensor(ctx, emission_grad_src, emission_grad_dst);
diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu
index 9431030a53975acafe9bcb22dc9164492929b07a..a3ab1a729761d59c058de1985b15575c9d50d3c5 100644
--- a/paddle/operators/lookup_table_op.cu
+++ b/paddle/operators/lookup_table_op.cu
@@ -101,7 +101,7 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       // copy GPU memory to CPU pinned memory
       framework::Vector<int64_t> new_rows;
       new_rows.resize(ids_dim[0]);
-      auto gpu_place = boost::get<platform::GPUPlace>(context.GetPlace());
+      auto gpu_place = boost::get<platform::CUDAPlace>(context.GetPlace());
 
       memory::Copy(platform::CPUPlace(), new_rows.data(), gpu_place, ids_data,
                    ids_dim[0] * sizeof(int64_t), stream);
diff --git a/paddle/operators/lstm_unit_op.cu b/paddle/operators/lstm_unit_op.cu
index 291f2c295e78288c01c6575df936ceedceba7ce8..4b164d964c3f56e52dc2b6ddcd945a241b39a7f9 100644
--- a/paddle/operators/lstm_unit_op.cu
+++ b/paddle/operators/lstm_unit_op.cu
@@ -98,7 +98,7 @@ class LstmUnitOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     auto* x_tensor = ctx.Input<framework::Tensor>("X");
     auto* c_prev_tensor = ctx.Input<framework::Tensor>("C_prev");
@@ -129,7 +129,7 @@ class LstmUnitGradOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     auto x_tensor = ctx.Input<Tensor>("X");
     auto c_prev_tensor = ctx.Input<Tensor>("C_prev");
diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc
index 256f3bc9bd487d11b0f139ef057f5a98556b4db1..26c038e435827b401d723ee6eef2255a89670f46 100644
--- a/paddle/operators/math/im2col_test.cc
+++ b/paddle/operators/math/im2col_test.cc
@@ -159,6 +159,7 @@ void testIm2col() {
 TEST(math, im2col) {
   testIm2col<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
 #ifdef PADDLE_WITH_CUDA
-  testIm2col<paddle::platform::CUDADeviceContext, paddle::platform::GPUPlace>();
+  testIm2col<paddle::platform::CUDADeviceContext,
+             paddle::platform::CUDAPlace>();
 #endif
 }
diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu
index 7852bb53a9035f71f52a51529c8e3cea22b0d4aa..0a818bc5d4575ac25f27c32b583780e1b6946f1e 100644
--- a/paddle/operators/math/math_function.cu
+++ b/paddle/operators/math/math_function.cu
@@ -105,7 +105,7 @@ void matmul<platform::CUDADeviceContext, float>(
   PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) &&
                      platform::is_gpu_place(matrix_b.place()) &&
                      platform::is_gpu_place(matrix_out->place()),
-                 "Matrix must all be in GPUPlace");
+                 "Matrix must all be in CUDAPlace");
 
   int M = dim_out[0];
   int N = dim_out[1];
@@ -134,7 +134,7 @@ void matmul<platform::CUDADeviceContext, double>(
   PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) &&
                      platform::is_gpu_place(matrix_b.place()) &&
                      platform::is_gpu_place(matrix_out->place()),
-                 "Matrix must all be in GPUPlace");
+                 "Matrix must all be in CUDAPlace");
 
   int M = dim_out[0];
   int N = dim_out[1];
@@ -266,7 +266,7 @@ struct TensorSetConstantGPU {
 };
 
 template <>
-void set_constant_with_place<platform::GPUPlace>(
+void set_constant_with_place<platform::CUDAPlace>(
     const platform::DeviceContext& context, framework::Tensor* tensor,
     float value) {
   framework::VisitDataType(framework::ToDataType(tensor->type()),
@@ -277,7 +277,7 @@ template <>
 void set_constant_with_place<platform::CUDNNPlace>(
     const platform::DeviceContext& context, framework::Tensor* tensor,
     float value) {
-  set_constant_with_place<platform::GPUPlace>(context, tensor, value);
+  set_constant_with_place<platform::CUDAPlace>(context, tensor, value);
 }
 
 template struct RowwiseAdd<platform::CUDADeviceContext, float>;
diff --git a/paddle/operators/math/math_function_test.cu b/paddle/operators/math/math_function_test.cu
index 32e96d948714a8fd1fa2c089057603fdaed85c16..4325a79664f15cfaea48870cd503ce70cc31044f 100644
--- a/paddle/operators/math/math_function_test.cu
+++ b/paddle/operators/math/math_function_test.cu
@@ -13,7 +13,7 @@ TEST(math_function, notrans_mul_trans) {
   float arr[6] = {0, 1, 2, 3, 4, 5};
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -47,7 +47,7 @@ TEST(math_function, trans_mul_notrans) {
   float arr[6] = {0, 1, 2, 3, 4, 5};
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -96,7 +96,7 @@ TEST(math_function, gemm_notrans_cublas) {
   float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -151,7 +151,7 @@ TEST(math_function, gemm_trans_cublas) {
   float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7};
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
   paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu);
@@ -189,7 +189,7 @@ void GemvTest(int m, int n, bool trans) {
   T* data_b = vec_b.mutable_data<T>({trans ? m : n}, *cpu_place);
   T* data_c = vec_c.mutable_data<T>({trans ? n : m}, *cpu_place);
 
-  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  auto* gpu_place = new paddle::platform::CUDAPlace(0);
   paddle::framework::Tensor g_mat_a;
   paddle::framework::Tensor g_vec_b;
   paddle::framework::Tensor g_vec_c;
diff --git a/paddle/operators/math/selected_rows_functor.cu b/paddle/operators/math/selected_rows_functor.cu
index c44577e00af5f362ae7e168495e496d60d05de95..9fddd97a36f7fdb6628d6eeb192cb216fdae3e5b 100644
--- a/paddle/operators/math/selected_rows_functor.cu
+++ b/paddle/operators/math/selected_rows_functor.cu
@@ -58,15 +58,15 @@ struct SelectedRowsAdd<platform::CUDADeviceContext, T> {
     PADDLE_ENFORCE(platform::is_gpu_place(out_place));
 
     memory::Copy(
-        boost::get<platform::GPUPlace>(out_place), out_data,
-        boost::get<platform::GPUPlace>(in1_place), in1_data,
+        boost::get<platform::CUDAPlace>(out_place), out_data,
+        boost::get<platform::CUDAPlace>(in1_place), in1_data,
         in1_value.numel() * sizeof(T),
         reinterpret_cast<const platform::CUDADeviceContext&>(context).stream());
 
     auto* in2_data = in2_value.data<T>();
-    memory::Copy(boost::get<platform::GPUPlace>(out_place),
+    memory::Copy(boost::get<platform::CUDAPlace>(out_place),
                  out_data + in1_value.numel(),
-                 boost::get<platform::GPUPlace>(in2_place), in2_data,
+                 boost::get<platform::CUDAPlace>(in2_place), in2_data,
                  in2_value.numel() * sizeof(T), context.stream());
   }
 };
@@ -160,9 +160,9 @@ struct SelectedRowsAddTo<platform::CUDADeviceContext, T> {
 
     auto* in1_data = in1_value.data<T>();
     auto* in2_data = in2_value->data<T>();
-    memory::Copy(boost::get<platform::GPUPlace>(in2_place),
+    memory::Copy(boost::get<platform::CUDAPlace>(in2_place),
                  in2_data + input2_offset,
-                 boost::get<platform::GPUPlace>(in1_place), in1_data,
+                 boost::get<platform::CUDAPlace>(in1_place), in1_data,
                  in1_value.numel() * sizeof(T), context.stream());
   }
 };
diff --git a/paddle/operators/math/selected_rows_functor_test.cu b/paddle/operators/math/selected_rows_functor_test.cu
index 777caf5635647d11e8fde05a68fdf7e2c32f48df..0a2e36f68acee04bd6b272d37679c18231cb8760 100644
--- a/paddle/operators/math/selected_rows_functor_test.cu
+++ b/paddle/operators/math/selected_rows_functor_test.cu
@@ -21,7 +21,7 @@ TEST(selected_rows_functor, gpu_add) {
   using namespace paddle::platform;
   using namespace paddle::operators::math;
 
-  GPUPlace gpu_place(0);
+  CUDAPlace gpu_place(0);
   CPUPlace cpu_place;
   CUDADeviceContext ctx(gpu_place);
   SetConstant<CUDADeviceContext, float> functor;
@@ -119,7 +119,7 @@ TEST(selected_rows_functor, gpu_add_to) {
   using namespace paddle::platform;
   using namespace paddle::operators::math;
 
-  GPUPlace gpu_place(0);
+  CUDAPlace gpu_place(0);
   CPUPlace cpu_place;
   CUDADeviceContext ctx(gpu_place);
   SetConstant<CUDADeviceContext, float> functor;
diff --git a/paddle/operators/math/vol2col_test.cc b/paddle/operators/math/vol2col_test.cc
index f46db3c56713399798a45854bf1613d07aee26e6..3794f0e52d200a08253a979991da04ec564cae47 100644
--- a/paddle/operators/math/vol2col_test.cc
+++ b/paddle/operators/math/vol2col_test.cc
@@ -122,6 +122,6 @@ TEST(math, vol2col) {
   testVol2col<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
 #ifdef PADDLE_WITH_CUDA
   testVol2col<paddle::platform::CUDADeviceContext,
-              paddle::platform::GPUPlace>();
+              paddle::platform::CUDAPlace>();
 #endif  // PADDLE_WITH_CUDA
 }
diff --git a/paddle/operators/multiplex_op.cu b/paddle/operators/multiplex_op.cu
index 47986e9ff86f2e08b0861cde35ac3a44b10caed1..57e6880b4e8b32eb36ce5d7f5c4a00ca0a3c48ff 100644
--- a/paddle/operators/multiplex_op.cu
+++ b/paddle/operators/multiplex_op.cu
@@ -36,7 +36,7 @@ class MultiplexGPUKernel : public framework::OpKernel<T> {
     CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
     auto stream = ctx.cuda_device_context().stream();
-    platform::GPUPlace place = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       int32_t k = index[i];
       PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative.");
@@ -73,7 +73,7 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
     auto* index = index_t_cpu.data<int32_t>();
 
     auto stream = ctx.cuda_device_context().stream();
-    platform::GPUPlace place = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       size_t k = static_cast<size_t>(index[i]);
       if (d_ins[k]) {
diff --git a/paddle/operators/nccl_op.cu.cc b/paddle/operators/nccl_op.cu.cc
index 6ca6db7253da0e59c742f115cd25a1b8203a3044..1b986a13650de7d77f4828d71798ee00d61c1284 100644
--- a/paddle/operators/nccl_op.cu.cc
+++ b/paddle/operators/nccl_op.cu.cc
@@ -67,7 +67,7 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
     auto stream = ctx.cuda_device_context().stream();
 
     // device id
-    int gpu_id = boost::get<platform::GPUPlace>(ctx.GetPlace()).GetDeviceId();
+    int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
 
     for (size_t i = 0; i < ins.size(); ++i) {
@@ -120,7 +120,7 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
                       ctx.device_context())
                       .stream();
     // device id
-    int gpu_id = boost::get<platform::GPUPlace>(ctx.GetPlace()).GetDeviceId();
+    int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
 
     auto ins_names = ctx.Inputs("X");
@@ -164,7 +164,7 @@ class NCCLBcastKernel : public framework::OpKernel<T> {
                       ctx.device_context())
                       .stream();
     // device id
-    int gpu_id = boost::get<platform::GPUPlace>(ctx.GetPlace()).GetDeviceId();
+    int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
 
     if (idx == root) {
diff --git a/paddle/operators/nccl_op_test.cu.cc b/paddle/operators/nccl_op_test.cu.cc
index b6e4ccb73f8133a55c17ddfa93b3ab4a21496561..361bfa8d75963bce7344a7263732f10190b078f3 100644
--- a/paddle/operators/nccl_op_test.cu.cc
+++ b/paddle/operators/nccl_op_test.cu.cc
@@ -52,7 +52,7 @@ class NCCLTester : public ::testing::Test {
   virtual void SetUp() override {
     paddle::platform::CPUPlace cpu_place;
     for (size_t i = 0; i < gpu_list.size(); ++i) {
-      p::GPUPlace place(i);
+      p::CUDAPlace place(i);
       dev_ctxs.emplace_back(new p::CUDADeviceContext(place));
     }
 
@@ -87,7 +87,7 @@ class NCCLTester : public ::testing::Test {
     std::unique_lock<std::mutex> lk(mu);
     const f::OpDesc *op1 = &op_desc;
 
-    p::GPUPlace place(gpu_id);
+    p::CUDAPlace place(gpu_id);
     auto &ctx = dev_ctxs.at(gpu_id);
 
     auto *send_tensor = scope->Var("st")->GetMutable<f::LoDTensor>();
@@ -171,7 +171,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) {
 
   for (size_t i = 0; i < dev_scopes.size(); ++i) {
     p::CPUPlace cpu_place;
-    p::GPUPlace gpu_place(gpu_list[i]);
+    p::CUDAPlace gpu_place(gpu_list[i]);
 
     auto &recv_tensor = dev_scopes[i]->FindVar("rt")->Get<f::LoDTensor>();
     auto *rt = recv_tensor.data<float>();
@@ -180,7 +180,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) {
     auto *ct = result_tensor->mutable_data<float>(cpu_place);
 
     paddle::memory::Copy(
-        cpu_place, ct, p::GPUPlace(gpu_list[i]), rt,
+        cpu_place, ct, p::CUDAPlace(gpu_list[i]), rt,
         recv_tensor.numel() * sizeof(float),
         static_cast<p::CUDADeviceContext *>(dev_ctxs[i])->stream());
 
@@ -219,7 +219,7 @@ TEST_F(NCCLTester, ncclReduceOp) {
   float result = std::accumulate(gpu_list.begin(), gpu_list.end(), 0);
 
   p::CPUPlace cpu_place;
-  p::GPUPlace gpu_place(gpu_list[kRoot]);
+  p::CUDAPlace gpu_place(gpu_list[kRoot]);
 
   auto &recv_tensor = dev_scopes[kRoot]->FindVar("rt")->Get<f::LoDTensor>();
   auto *rt = recv_tensor.data<float>();
@@ -229,7 +229,7 @@ TEST_F(NCCLTester, ncclReduceOp) {
   auto *ct = result_tensor->mutable_data<float>(cpu_place);
 
   paddle::memory::Copy(
-      cpu_place, ct, p::GPUPlace(gpu_list[kRoot]), rt,
+      cpu_place, ct, p::CUDAPlace(gpu_list[kRoot]), rt,
       recv_tensor.numel() * sizeof(float),
       static_cast<p::CUDADeviceContext *>(dev_ctxs[kRoot])->stream());
 
@@ -268,7 +268,7 @@ TEST_F(NCCLTester, ncclBcastOp) {
   float result = kRoot;
 
   p::CPUPlace cpu_place;
-  p::GPUPlace gpu_place(gpu_list[idx]);
+  p::CUDAPlace gpu_place(gpu_list[idx]);
 
   auto &recv_tensor = dev_scopes[idx]->FindVar("rt")->Get<f::LoDTensor>();
   auto *rt = recv_tensor.data<float>();
@@ -277,7 +277,7 @@ TEST_F(NCCLTester, ncclBcastOp) {
   auto *ct = result_tensor->mutable_data<float>(cpu_place);
 
   paddle::memory::Copy(
-      cpu_place, ct, p::GPUPlace(gpu_list[idx]), rt,
+      cpu_place, ct, p::CUDAPlace(gpu_list[idx]), rt,
       recv_tensor.numel() * sizeof(float),
       static_cast<p::CUDADeviceContext *>(dev_ctxs[idx])->stream());
 
@@ -300,7 +300,7 @@ int main(int argc, char **argv) {
   places.emplace_back(paddle::platform::CPUPlace());
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; ++i) {
-    places.emplace_back(paddle::platform::GPUPlace(i));
+    places.emplace_back(paddle::platform::CUDAPlace(i));
     gpu_list.emplace_back(i);
   }
 
diff --git a/paddle/operators/pool_cudnn_op.cu.cc b/paddle/operators/pool_cudnn_op.cu.cc
index fc2b37bd0fbac82005e709779b2939843b839596..2d0001ba1184c99d9fc642f60c97ba89cec97ccd 100644
--- a/paddle/operators/pool_cudnn_op.cu.cc
+++ b/paddle/operators/pool_cudnn_op.cu.cc
@@ -29,7 +29,7 @@ class PoolCudnnOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     const Tensor *input = ctx.Input<Tensor>("X");
     Tensor *output = ctx.Output<Tensor>("Out");
@@ -90,7 +90,7 @@ class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
 
     const Tensor *input = ctx.Input<Tensor>("X");
     const Tensor *output = ctx.Input<Tensor>("Out");
diff --git a/paddle/operators/reshape_op.cu b/paddle/operators/reshape_op.cu
index b7329238c0ea8ebb374d35bd7cddced3dfee1a2c..a5dcd2ec9647eb3085ea8cbb2f9edbff9ab9cf92 100644
--- a/paddle/operators/reshape_op.cu
+++ b/paddle/operators/reshape_op.cu
@@ -16,7 +16,7 @@
 
 REGISTER_OP_CUDA_KERNEL(
     reshape,
-    paddle::operators::ReshapeKernel<paddle::platform::GPUPlace, float>);
+    paddle::operators::ReshapeKernel<paddle::platform::CUDAPlace, float>);
 REGISTER_OP_CUDA_KERNEL(
     reshape_grad,
-    paddle::operators::ReshapeGradKernel<paddle::platform::GPUPlace, float>);
+    paddle::operators::ReshapeGradKernel<paddle::platform::CUDAPlace, float>);
diff --git a/paddle/operators/strided_memcpy_test.cc b/paddle/operators/strided_memcpy_test.cc
index 230cc1ab0bbd5ac57eb7494795e3fbcdf02c3cc8..d47fd98d06e52b55056521d08231fa9316289a2b 100644
--- a/paddle/operators/strided_memcpy_test.cc
+++ b/paddle/operators/strided_memcpy_test.cc
@@ -82,7 +82,7 @@ TEST(StridedMemcpy, GPUCrop) {
   };
   // clang-format on
 
-  platform::GPUPlace gpu0(0);
+  platform::CUDAPlace gpu0(0);
   platform::CPUPlace cpu;
 
   platform::CUDADeviceContext ctx(gpu0);
@@ -121,7 +121,7 @@ TEST(StridedMemcpy, GPUConcat) {
   };
   // clang-format on
 
-  platform::GPUPlace gpu0(0);
+  platform::CUDAPlace gpu0(0);
   platform::CPUPlace cpu;
   platform::CUDADeviceContext ctx(gpu0);
 
diff --git a/paddle/operators/top_k_op.cu b/paddle/operators/top_k_op.cu
index 453bd07267e3a6e33211117368dd9aff10a9e23f..0a70ad87e672080f11439f44f0c9bfcf01114fd6 100644
--- a/paddle/operators/top_k_op.cu
+++ b/paddle/operators/top_k_op.cu
@@ -283,7 +283,7 @@ class TopkOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
+                   "It must use CUDAPlace.");
     auto* input = ctx.Input<Tensor>("X");
     auto* output = ctx.Output<Tensor>("Out");
     auto* indices = ctx.Output<Tensor>("Indices");
diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc
index a28e9de716c857145955cced85b99b77ef89b101..8ee0f18e64599869670bd6665df43027cfde9501 100644
--- a/paddle/platform/device_context.cc
+++ b/paddle/platform/device_context.cc
@@ -58,10 +58,10 @@ DeviceContextPool::DeviceContextPool(
 #ifdef PADDLE_WITH_CUDA
       device_contexts_.emplace(places[i],
                                new platform::CUDADeviceContext(
-                                   boost::get<platform::GPUPlace>(places[i])));
+                                   boost::get<platform::CUDAPlace>(places[i])));
 #else
       PADDLE_THROW(
-          "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
+          "'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
           "option");
 #endif
     }
@@ -91,7 +91,7 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface {
   }
   ~EigenCudaStreamDevice() override {}
 
-  void Reinitialize(const cudaStream_t* cuda_stream, GPUPlace place) {
+  void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) {
     stream_ = cuda_stream;
     place_ = place;
     device_prop_ = &Eigen::m_deviceProperties[place.device];
@@ -130,14 +130,14 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface {
   }
 
  private:
-  GPUPlace place_;
+  CUDAPlace place_;
   const cudaStream_t* stream_;         // not owned;
   const cudaDeviceProp* device_prop_;  // not owned;
   mutable void* scratch_;
   mutable unsigned int* semaphore_;
 };
 
-CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) {
+CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) {
   SetDeviceId(place_.device);
   PADDLE_ENFORCE(cudaStreamCreate(&stream_));
   eigen_stream_.reset(new EigenCudaStreamDevice());
diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h
index 9b958f7c920a32c9208f3dfd3ff54ac9620da9e7..877a66363a1a523692b07e7bef0e11e3caddce9f 100644
--- a/paddle/platform/device_context.h
+++ b/paddle/platform/device_context.h
@@ -58,7 +58,7 @@ class EigenCudaStreamDevice;
 
 class CUDADeviceContext : public DeviceContext {
  public:
-  explicit CUDADeviceContext(GPUPlace place);
+  explicit CUDADeviceContext(CUDAPlace place);
   virtual ~CUDADeviceContext();
 
   /*! \brief  Wait for all operations completion in the stream. */
@@ -80,7 +80,7 @@ class CUDADeviceContext : public DeviceContext {
   cudaStream_t stream() const;
 
  private:
-  GPUPlace place_;
+  CUDAPlace place_;
 
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
@@ -143,7 +143,7 @@ class DeviceContextPool {
     size_t operator()(const platform::Place& place) const {
       int pre_hash = place.which() + (1 << LEFT_SHIFT);
       if (platform::is_gpu_place(place)) {
-        pre_hash += boost::get<platform::GPUPlace>(place).GetDeviceId();
+        pre_hash += boost::get<platform::CUDAPlace>(place).GetDeviceId();
       }
       return hash_(pre_hash);
     }
diff --git a/paddle/platform/device_context_test.cu b/paddle/platform/device_context_test.cu
index f046c79e0a015023568071a157ae183bfb8df556..186824c0194a2e1d300e60d1e8a84b2f65a20988 100644
--- a/paddle/platform/device_context_test.cu
+++ b/paddle/platform/device_context_test.cu
@@ -20,11 +20,11 @@ limitations under the License. */
 TEST(Device, Init) {
   using paddle::platform::DeviceContext;
   using paddle::platform::CUDADeviceContext;
-  using paddle::platform::GPUPlace;
+  using paddle::platform::CUDAPlace;
 
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; i++) {
-    CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
+    CUDADeviceContext* device_context = new CUDADeviceContext(CUDAPlace(i));
     Eigen::GpuDevice* gpu_device = device_context->eigen_device();
     ASSERT_NE(nullptr, gpu_device);
     delete device_context;
@@ -33,11 +33,11 @@ TEST(Device, Init) {
 
 TEST(Device, CUDADeviceContext) {
   using paddle::platform::CUDADeviceContext;
-  using paddle::platform::GPUPlace;
+  using paddle::platform::CUDAPlace;
 
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; i++) {
-    CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
+    CUDADeviceContext* device_context = new CUDADeviceContext(CUDAPlace(i));
     Eigen::GpuDevice* gpu_device = device_context->eigen_device();
     ASSERT_NE(nullptr, gpu_device);
     cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
@@ -70,7 +70,7 @@ TEST(Device, DeviceContextPool) {
   using paddle::platform::CUDADeviceContext;
   using paddle::platform::Place;
   using paddle::platform::CPUPlace;
-  using paddle::platform::GPUPlace;
+  using paddle::platform::CUDAPlace;
 
   DeviceContextPool& pool = DeviceContextPool::Get();
   auto cpu_dev_ctx1 = pool.Borrow(CPUPlace());
@@ -80,14 +80,14 @@ TEST(Device, DeviceContextPool) {
   std::vector<Place> gpu_places;
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; ++i) {
-    gpu_places.emplace_back(GPUPlace(i));
+    gpu_places.emplace_back(CUDAPlace(i));
   }
   auto dev_ctxs = pool.Borrow(gpu_places);
   for (size_t i = 0; i < dev_ctxs.size(); ++i) {
     auto* dev_ctx = static_cast<const CUDADeviceContext*>(dev_ctxs[i]);
 
-    // check same as GPUPlace(i)
-    GPUPlace place = boost::get<GPUPlace>(dev_ctx->GetPlace());
+    // check same as CUDAPlace(i)
+    CUDAPlace place = boost::get<CUDAPlace>(dev_ctx->GetPlace());
     EXPECT_EQ(place.GetDeviceId(), static_cast<int>(i));
   }
 }
@@ -106,7 +106,7 @@ int main(int argc, char** argv) {
   places.emplace_back(paddle::platform::CPUPlace());
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; ++i) {
-    places.emplace_back(paddle::platform::GPUPlace(i));
+    places.emplace_back(paddle::platform::CUDAPlace(i));
   }
 
   VLOG(0) << " DeviceCount " << count;
diff --git a/paddle/platform/nccl_test.cu b/paddle/platform/nccl_test.cu
index 6750c8da7db86500e9593cd41d39dbd229abad7a..f57c329402da4854b2028bf5836e5d7a013bf8e2 100644
--- a/paddle/platform/nccl_test.cu
+++ b/paddle/platform/nccl_test.cu
@@ -50,7 +50,7 @@ struct PerThreadData {
 
   T* RecvBuff() { return thrust::raw_pointer_cast(recv_buff.data()); }
 
-  PerThreadData(int gpu_id, size_t size) : dev_ctx(GPUPlace(gpu_id)) {
+  PerThreadData(int gpu_id, size_t size) : dev_ctx(CUDAPlace(gpu_id)) {
     send_buff.resize(size);
     for (size_t i = 0; i < size; ++i) {
       send_buff[i] = static_cast<T>(i);
@@ -140,7 +140,7 @@ int main(int argc, char** argv) {
   places.emplace_back(paddle::platform::CPUPlace());
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; ++i) {
-    places.emplace_back(paddle::platform::GPUPlace(i));
+    places.emplace_back(paddle::platform::CUDAPlace(i));
   }
 
   VLOG(0) << " DeviceCount " << count;
diff --git a/paddle/platform/place.cc b/paddle/platform/place.cc
index 25fe8d21b49b07a6afe2938245906dc1bdd90398..4d23cfd8865bd83e9702169e3b3d5f88b8a32059 100644
--- a/paddle/platform/place.cc
+++ b/paddle/platform/place.cc
@@ -24,7 +24,9 @@ class PlacePrinter : public boost::static_visitor<> {
   explicit PlacePrinter(std::ostream &os) : os_(os) {}
   void operator()(const CPUPlace &) { os_ << "CPUPlace"; }
   void operator()(const MKLDNNPlace &) { os_ << "MKLDNNPlace"; }
-  void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; }
+  void operator()(const CUDAPlace &p) {
+    os_ << "CUDAPlace(" << p.device << ")";
+  }
 
  private:
   std::ostream &os_;
@@ -37,12 +39,12 @@ static Place the_default_place;
 void set_place(const Place &place) { the_default_place = place; }
 const Place &get_place() { return the_default_place; }
 
-const GPUPlace default_gpu() { return GPUPlace(0); }
+const CUDAPlace default_gpu() { return CUDAPlace(0); }
 const CPUPlace default_cpu() { return CPUPlace(); }
 const MKLDNNPlace default_mkldnn() { return MKLDNNPlace(); }
 
 bool is_gpu_place(const Place &p) {
-  return boost::apply_visitor(IsGPUPlace(), p);
+  return boost::apply_visitor(IsCUDAPlace(), p);
 }
 bool is_cpu_place(const Place &p) {
   return !is_gpu_place(p) && !is_mkldnn_place(p);
diff --git a/paddle/platform/place.h b/paddle/platform/place.h
index daeafbbcd780aaeab20c8fcbbeed60a587e0049b..4eab1a3964bddcce40412450263610b7e0824721 100644
--- a/paddle/platform/place.h
+++ b/paddle/platform/place.h
@@ -39,43 +39,45 @@ struct MKLDNNPlace {
   inline bool operator!=(const MKLDNNPlace &) const { return false; }
 };
 
-struct GPUPlace {
-  GPUPlace() : GPUPlace(0) {}
-  explicit GPUPlace(int d) : device(d) {}
+struct CUDAPlace {
+  CUDAPlace() : CUDAPlace(0) {}
+  explicit CUDAPlace(int d) : device(d) {}
 
   inline int GetDeviceId() const { return device; }
   // needed for variant equality comparison
-  inline bool operator==(const GPUPlace &o) const { return device == o.device; }
-  inline bool operator!=(const GPUPlace &o) const { return !(*this == o); }
+  inline bool operator==(const CUDAPlace &o) const {
+    return device == o.device;
+  }
+  inline bool operator!=(const CUDAPlace &o) const { return !(*this == o); }
 
   int device;
 };
 
-struct CUDNNPlace : public GPUPlace {
-  CUDNNPlace() : GPUPlace() {}
-  explicit CUDNNPlace(int d) : GPUPlace(d) {}
+struct CUDNNPlace : public CUDAPlace {
+  CUDNNPlace() : CUDAPlace() {}
+  explicit CUDNNPlace(int d) : CUDAPlace(d) {}
 };
 
-struct IsGPUPlace : public boost::static_visitor<bool> {
+struct IsCUDAPlace : public boost::static_visitor<bool> {
   bool operator()(const CPUPlace &) const { return false; }
   bool operator()(const MKLDNNPlace &) const { return false; }
-  bool operator()(const GPUPlace &gpu) const { return true; }
+  bool operator()(const CUDAPlace &gpu) const { return true; }
   bool operator()(const CUDNNPlace &) const { return true; }
 };
 
 struct IsMKLDNNPlace : public boost::static_visitor<bool> {
   bool operator()(const MKLDNNPlace &) const { return true; }
   bool operator()(const CPUPlace &) const { return false; }
-  bool operator()(const GPUPlace &) const { return false; }
+  bool operator()(const CUDAPlace &) const { return false; }
   bool operator()(const CUDNNPlace &) const { return false; }
 };
 
-typedef boost::variant<CUDNNPlace, GPUPlace, CPUPlace, MKLDNNPlace> Place;
+typedef boost::variant<CUDNNPlace, CUDAPlace, CPUPlace, MKLDNNPlace> Place;
 
 void set_place(const Place &);
 const Place &get_place();
 
-const GPUPlace default_gpu();
+const CUDAPlace default_gpu();
 const CPUPlace default_cpu();
 const MKLDNNPlace default_mkldnn();
 
diff --git a/paddle/platform/place_test.cc b/paddle/platform/place_test.cc
index c536b59ed8f71bd078bd09c5bd5afeab74c71b28..21f7d9f21323ec6d494c3c3a2fa53d07a31e9edd 100644
--- a/paddle/platform/place_test.cc
+++ b/paddle/platform/place_test.cc
@@ -4,7 +4,7 @@
 
 TEST(Place, Equality) {
   paddle::platform::CPUPlace cpu;
-  paddle::platform::GPUPlace g0(0), g1(1), gg0(0);
+  paddle::platform::CUDAPlace g0(0), g1(1), gg0(0);
   paddle::platform::CUDNNPlace d0(0), d1(1), dd0(0);
 
   EXPECT_EQ(cpu, cpu);
@@ -41,8 +41,8 @@ TEST(Place, Default) {
 TEST(Place, Print) {
   {
     std::stringstream ss;
-    ss << paddle::platform::GPUPlace(1);
-    EXPECT_EQ("GPUPlace(1)", ss.str());
+    ss << paddle::platform::CUDAPlace(1);
+    EXPECT_EQ("CUDAPlace(1)", ss.str());
   }
   {
     std::stringstream ss;
diff --git a/paddle/platform/transform_test.cu b/paddle/platform/transform_test.cu
index 464096111e4a85b8d64d9223bfb85a1d1d42fad4..8e2483aa84568a6986a923670f71050a12b55136 100644
--- a/paddle/platform/transform_test.cu
+++ b/paddle/platform/transform_test.cu
@@ -49,7 +49,7 @@ TEST(Transform, CPUUnary) {
 TEST(Transform, GPUUnary) {
   using namespace paddle::platform;
   using namespace paddle::memory;
-  GPUPlace gpu0(0);
+  CUDAPlace gpu0(0);
   CUDADeviceContext ctx(gpu0);
   float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
   float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
@@ -80,7 +80,7 @@ TEST(Transform, GPUBinary) {
   using namespace paddle::platform;
   using namespace paddle::memory;
   int buf[4] = {1, 2, 3, 4};
-  GPUPlace gpu0(0);
+  CUDAPlace gpu0(0);
   CUDADeviceContext ctx(gpu0);
   int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
   Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf), ctx.stream());
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index de6b24f70d84a28add0c0a09cac79b8c5b1044de..668a48e816919cbc6a4bb646adedfbbdcb57922a 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -79,7 +79,7 @@ PYBIND11_PLUGIN(core) {
              self.Resize(make_ddim(dim));
            })
       .def("alloc_float",
-           [](Tensor &self, paddle::platform::GPUPlace &place) {
+           [](Tensor &self, paddle::platform::CUDAPlace &place) {
              self.mutable_data<float>(place);
            })
       .def("alloc_float",
@@ -91,7 +91,7 @@ PYBIND11_PLUGIN(core) {
              self.mutable_data<int>(place);
            })
       .def("alloc_int",
-           [](Tensor &self, paddle::platform::GPUPlace &place) {
+           [](Tensor &self, paddle::platform::CUDAPlace &place) {
              self.mutable_data<int>(place);
            })
       .def("set", PyCPUTensorSetFromArray<float>)
@@ -310,10 +310,10 @@ All parameter, weight, gradient are variables in Paddle.
                     return new paddle::platform::CPUDeviceContext();
                   })
       .def_static("create",
-                  [](paddle::platform::GPUPlace& place)
+                  [](paddle::platform::CUDAPlace& place)
                       -> paddle::platform::DeviceContext* {
 #ifndef PADDLE_WITH_CUDA
-                    PADDLE_THROW("GPUPlace is not supported in CPU device.");
+                    PADDLE_THROW("CUDAPlace is not supported in CPU device.");
 #else
                     return new paddle::platform::CUDADeviceContext(place);
 #endif
@@ -323,9 +323,9 @@ All parameter, weight, gradient are variables in Paddle.
 #ifdef PADDLE_WITH_CUDA
   py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
 #endif
-  py::class_<platform::GPUPlace>(m, "GPUPlace")
+  py::class_<platform::CUDAPlace>(m, "CUDAPlace")
       .def(py::init<int>())
-      .def("__str__", string::to_string<const platform::GPUPlace &>);
+      .def("__str__", string::to_string<const platform::CUDAPlace &>);
 
   py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
       .def(py::init<>())
@@ -338,7 +338,7 @@ All parameter, weight, gradient are variables in Paddle.
              self = cpu_place;
            })
       .def("set_place",
-           [](platform::Place &self, const platform::GPUPlace &gpu_place) {
+           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
              self = gpu_place;
            });
 
@@ -363,7 +363,7 @@ All parameter, weight, gradient are variables in Paddle.
               const platform::CPUPlace &place) { self.Run(scope, place); })
       .def("run",
            [](OperatorBase &self, const Scope &scope,
-              const platform::GPUPlace &place) { self.Run(scope, place); })
+              const platform::CUDAPlace &place) { self.Run(scope, place); })
       .def("type",
            [](const OperatorBase &op) -> std::string { return op.Type(); })
       .def("outputs",
diff --git a/paddle/pybind/tensor_py.h b/paddle/pybind/tensor_py.h
index 413fd9b046f3f302feb5bd52beb284553a8ae192..7b8c29ff84ff4950a1ecf2d6364793b8567bfd4e 100644
--- a/paddle/pybind/tensor_py.h
+++ b/paddle/pybind/tensor_py.h
@@ -71,7 +71,7 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
             dst_ptr, src_ptr, sizeof(CUR_TYPE) * tensor.numel(),
             cudaMemcpyDeviceToHost, dev_ctx->stream());
 #else
-        PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
+        PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
 #endif
       } else if (paddle::platform::is_cpu_place(tensor.place())) {
         dst_tensor = tensor;
@@ -127,7 +127,7 @@ template <typename T>
 void PyCUDATensorSetFromArray(
     framework::Tensor &self,
     py::array_t<T, py::array::c_style | py::array::forcecast> array,
-    paddle::platform::GPUPlace &place) {
+    paddle::platform::CUDAPlace &place) {
   std::vector<int64_t> dims;
   dims.reserve(array.ndim());
   for (size_t i = 0; i < array.ndim(); ++i) {
diff --git a/paddle/testing/paddle_gtest_main.cc b/paddle/testing/paddle_gtest_main.cc
index 7ba1bf095ab74e4b64a8fb39b84172d6f371a2cf..108ff335bf6b920c648d4bfebbd6a40ffb6fd939 100644
--- a/paddle/testing/paddle_gtest_main.cc
+++ b/paddle/testing/paddle_gtest_main.cc
@@ -36,7 +36,7 @@ int main(int argc, char** argv) {
   paddle::memory::Used(paddle::platform::CPUPlace());
   std::vector<std::string> devs = {"CPU"};
 #ifdef PADDLE_WITH_CUDA
-  paddle::memory::Used(paddle::platform::GPUPlace(0));
+  paddle::memory::Used(paddle::platform::CUDAPlace(0));
   devs.push_back("GPU:0");
 #endif
   paddle::framework::InitDevices(devs);
diff --git a/python/paddle/v2/fluid/__init__.py b/python/paddle/v2/fluid/__init__.py
index 051b9094aafa74b186776ae2041f95d0fe6d5f77..c72b5730695dbc4f772015f1fb8dec6814cd1837 100644
--- a/python/paddle/v2/fluid/__init__.py
+++ b/python/paddle/v2/fluid/__init__.py
@@ -15,14 +15,14 @@ import backward
 import regularizer
 from param_attr import ParamAttr
 from data_feeder import DataFeeder
-from core import LoDTensor, CPUPlace, GPUPlace
+from core import LoDTensor, CPUPlace, CUDAPlace
 from distribute_transpiler import DistributeTranspiler
 import clip
 
 Tensor = LoDTensor
 __all__ = framework.__all__ + executor.__all__ + [
     'io', 'initializer', 'layers', 'nets', 'optimizer', 'backward',
-    'regularizer', 'LoDTensor', 'CPUPlace', 'GPUPlace', 'Tensor', 'ParamAttr'
+    'regularizer', 'LoDTensor', 'CPUPlace', 'CUDAPlace', 'Tensor', 'ParamAttr'
     'DataFeeder', 'clip', 'DistributeTranspiler'
 ]
 
diff --git a/python/paddle/v2/fluid/executor.py b/python/paddle/v2/fluid/executor.py
index cdd576294f4f53bd3760b2c95a41b2129004a51a..2c91afb363bf72f2791e60c6df0d9130ccd698c5 100644
--- a/python/paddle/v2/fluid/executor.py
+++ b/python/paddle/v2/fluid/executor.py
@@ -47,7 +47,7 @@ class Executor(object):
             act_places.append(p)
 
         # TODO(dzhwinter) : consider that our fluid tests all written in 
-        # GPUPlace(gpu_id), this will be changed in the future
+        # CUDAPlace(gpu_id), this will be changed in the future
         if core.is_compile_gpu():
             core.init_devices(["CPU", "GPU:0"])
         else:
diff --git a/python/paddle/v2/fluid/tests/book/test_recommender_system.py b/python/paddle/v2/fluid/tests/book/test_recommender_system.py
index b0c11ba3414d6d87d078189ccee0791921040d91..e3cc2a89371233014dec4ba3d730a866722d3eae 100644
--- a/python/paddle/v2/fluid/tests/book/test_recommender_system.py
+++ b/python/paddle/v2/fluid/tests/book/test_recommender_system.py
@@ -142,7 +142,7 @@ def main():
     opts = sgd_optimizer.minimize(cost)
 
     if USE_GPU:
-        place = core.GPUPlace(0)
+        place = core.CUDAPlace(0)
     else:
         place = core.CPUPlace()
 
diff --git a/python/paddle/v2/fluid/tests/op_test.py b/python/paddle/v2/fluid/tests/op_test.py
index 087283bfded07e25ddfd446849b9c5ca9d1e7651..8dbfbd547a6677517f028997e6269709aac43b67 100644
--- a/python/paddle/v2/fluid/tests/op_test.py
+++ b/python/paddle/v2/fluid/tests/op_test.py
@@ -316,7 +316,7 @@ class OpTest(unittest.TestCase):
     def check_output(self, atol=1e-5):
         places = [core.CPUPlace()]
         if core.is_compile_gpu() and core.op_support_gpu(self.op_type):
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
         for place in places:
             self.check_output_with_place(place, atol)
 
@@ -379,7 +379,7 @@ class OpTest(unittest.TestCase):
                                "Gradient Check On %s" % str(cpu_place))
 
         if core.is_compile_gpu() and self.op.support_gpu():
-            gpu_place = core.GPUPlace(0)
+            gpu_place = core.CUDAPlace(0)
             gpu_analytic_grads = self._get_gradient(inputs_to_check, gpu_place,
                                                     output_names, no_grad_set)
 
diff --git a/python/paddle/v2/fluid/tests/test_adagrad_op.py b/python/paddle/v2/fluid/tests/test_adagrad_op.py
index 1ff3932164bed75be71b5c6b7114df362b893f09..7b2d02fbf4256d2c27383a3452d526271af543a3 100644
--- a/python/paddle/v2/fluid/tests/test_adagrad_op.py
+++ b/python/paddle/v2/fluid/tests/test_adagrad_op.py
@@ -167,7 +167,7 @@ class TestSparseAdagradOp(unittest.TestCase):
     def test_sparse_adagrad(self):
         places = [core.CPUPlace()]
         if core.is_compile_gpu():
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
         for place in places:
             self.check_with_place(place)
 
diff --git a/python/paddle/v2/fluid/tests/test_batch_norm_op.py b/python/paddle/v2/fluid/tests/test_batch_norm_op.py
index dfc047e1f0dc9fcf3d72d007b17d4c2de2077fbd..abbd48d2b843cedb77caffc13413d2f9695defa6 100644
--- a/python/paddle/v2/fluid/tests/test_batch_norm_op.py
+++ b/python/paddle/v2/fluid/tests/test_batch_norm_op.py
@@ -304,7 +304,7 @@ class TestBatchNormOp(OpTest):
             self.__assert_close(saved_variance_tensor, saved_variance,
                                 "saved_variance")
             self.__assert_close(mean_out_tensor, mean_out, "mean_out")
-            if isinstance(place, core.GPUPlace):
+            if isinstance(place, core.CUDAPlace):
                 atol = 5e-2
             else:
                 atol = 1e-4
@@ -339,7 +339,7 @@ class TestBatchNormOp(OpTest):
 
         places = [core.CPUPlace()]
         if core.is_compile_gpu() and core.op_support_gpu("batch_norm"):
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
 
             core.init_devices(["CPU", "GPU:0"])
         else:
diff --git a/python/paddle/v2/fluid/tests/test_gaussian_random_op.py b/python/paddle/v2/fluid/tests/test_gaussian_random_op.py
index 4afe0c6a6d36b4ab7b88b459ce8d182b287b860e..6f6a60ccb3ff17f6a12eec6974b8b2d73885c29f 100644
--- a/python/paddle/v2/fluid/tests/test_gaussian_random_op.py
+++ b/python/paddle/v2/fluid/tests/test_gaussian_random_op.py
@@ -20,7 +20,7 @@ class TestGaussianRandomOp(unittest.TestCase):
 
     def test_gpu(self):
         if core.is_compile_gpu():
-            self.gaussian_random_test(place=fluid.GPUPlace(0))
+            self.gaussian_random_test(place=fluid.CUDAPlace(0))
 
     def gaussian_random_test(self, place):
 
diff --git a/python/paddle/v2/fluid/tests/test_profiler.py b/python/paddle/v2/fluid/tests/test_profiler.py
index d01e257449542ad60741af29dff7328e8c1306b6..e3f3ac58ef9b30864849770510f7339749dab84f 100644
--- a/python/paddle/v2/fluid/tests/test_profiler.py
+++ b/python/paddle/v2/fluid/tests/test_profiler.py
@@ -15,7 +15,7 @@ class TestProfiler(unittest.TestCase):
         data = layers.data(name='data', shape=[3, 28, 28], dtype='float32')
         conv = layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1])
 
-        place = fluid.GPUPlace(0)
+        place = fluid.CUDAPlace(0)
         exe = fluid.Executor(place)
         exe.run(fluid.default_startup_program())
 
diff --git a/python/paddle/v2/fluid/tests/test_sgd_op.py b/python/paddle/v2/fluid/tests/test_sgd_op.py
index 9c345792beef46f65ec12e111f1d645fb31e69c7..14d41e172a22c677235ab3fa997ef6f0b6e39778 100644
--- a/python/paddle/v2/fluid/tests/test_sgd_op.py
+++ b/python/paddle/v2/fluid/tests/test_sgd_op.py
@@ -78,7 +78,7 @@ class TestSparseSGDOp(unittest.TestCase):
     def test_sparse_sgd(self):
         places = [core.CPUPlace()]
         if core.is_compile_gpu():
-            places.append(core.GPUPlace(0))
+            places.append(core.CUDAPlace(0))
         for place in places:
             self.check_with_place(place)
 
diff --git a/python/paddle/v2/fluid/tests/test_uniform_random_op.py b/python/paddle/v2/fluid/tests/test_uniform_random_op.py
index d6872c8ba351a13b6fb8622cc23029c8c5cbe2e1..dbe4d6bcd069d2088b3cc1b4efd575d14afd4198 100644
--- a/python/paddle/v2/fluid/tests/test_uniform_random_op.py
+++ b/python/paddle/v2/fluid/tests/test_uniform_random_op.py
@@ -23,7 +23,7 @@ class TestUniformRandomOp(unittest.TestCase):
 
     def test_gpu(self):
         if core.is_compile_gpu():
-            self.uniform_random_test(place=core.GPUPlace(0))
+            self.uniform_random_test(place=core.CUDAPlace(0))
 
     def uniform_random_test(self, place):
         program = fluid.Program()