diff --git a/paddle/fluid/operators/conv_cudnn_op_cache.h b/paddle/fluid/operators/conv_cudnn_op_cache.h
index ddddb7f864..23a471cfa0 100644
--- a/paddle/fluid/operators/conv_cudnn_op_cache.h
+++ b/paddle/fluid/operators/conv_cudnn_op_cache.h
@@ -40,11 +40,6 @@ static constexpr size_t kNUM_CUDNN_BWD_FILTER_ALGS =
     CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT;
 static constexpr size_t kNUM_CUDNN_BWD_DATA_ALGS =
     CUDNN_CONVOLUTION_BWD_DATA_ALGO_COUNT;
-#else
-// cuDNN v5 has no CUDNN_CONVOLUTION_FWD_ALGO_COUNT etc.
-static constexpr size_t kNUM_CUDNN_FWD_ALGS = 7;
-static constexpr size_t kNUM_CUDNN_BWD_FILTER_ALGS = 4;
-static constexpr size_t kNUM_CUDNN_BWD_DATA_ALGS = 5;
 #endif
 
 }  // namespace operators
diff --git a/paddle/fluid/operators/cudnn_lstm_cache.h b/paddle/fluid/operators/cudnn_lstm_cache.h
index 3181e4b1d9..b7859237e7 100644
--- a/paddle/fluid/operators/cudnn_lstm_cache.h
+++ b/paddle/fluid/operators/cudnn_lstm_cache.h
@@ -85,20 +85,12 @@ class ScopedRNNBase {
     dropout_desc_.descriptor(handle, place, initialized_, dropout_prob_,
                              dropout_state, seed_, state_size);
 
-// ------------------- cudnn rnn descriptors ---------------------
-#if CUDNN_VERSION >= 6000
+    // ------------------- cudnn rnn descriptors ---------------------
     PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetRNNDescriptor_v6(
         handle, rnn_desc_.desc(), hidden_size_, num_layers_,
         dropout_desc_.desc(), CUDNN_LINEAR_INPUT,
         is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM,
         CUDNN_RNN_ALGO_STANDARD, cudnn_type));
-#else
-    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetRNNDescriptor(
-        rnn_desc_.desc(), hidden_size_, num_layers_, dropout_desc_.desc(),
-        CUDNN_LINEAR_INPUT,
-        is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM,
-        cudnn_type));
-#endif
 
 #if CUDNN_VERSION >= 7201
     if (!sequence_length.empty()) {
diff --git a/paddle/fluid/operators/cudnn_rnn_cache.h b/paddle/fluid/operators/cudnn_rnn_cache.h
index 13a3e7d09b..a6a23a91c7 100644
--- a/paddle/fluid/operators/cudnn_rnn_cache.h
+++ b/paddle/fluid/operators/cudnn_rnn_cache.h
@@ -168,18 +168,11 @@ struct CudnnRNNCache {
     PADDLE_ENFORCE_CUDA_SUCCESS(
         platform::dynload::cudnnCreateRNNDescriptor(&rnn_desc_));
 
-#if CUDNN_VERSION >= 6000
     PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetRNNDescriptor_v6(
         handle, rnn_desc_, hidden_size_, num_layers_, dropout_desc_,
         CUDNN_LINEAR_INPUT,
         is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM,
         CUDNN_RNN_ALGO_STANDARD, cudnn_type));
-#else
-    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetRNNDescriptor(
-        rnn_desc_, hidden_size_, num_layers_, dropout_desc_, CUDNN_LINEAR_INPUT,
-        is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM,
-        cudnn_type));
-#endif
 
     PADDLE_ENFORCE_CUDA_SUCCESS(
         platform::dynload::cudnnCreateFilterDescriptor(&w_desc_));
diff --git a/paddle/fluid/platform/cudnn_helper.h b/paddle/fluid/platform/cudnn_helper.h
index af0df2efc5..6c3c96b68c 100644
--- a/paddle/fluid/platform/cudnn_helper.h
+++ b/paddle/fluid/platform/cudnn_helper.h
@@ -91,30 +91,6 @@ enum class ActivationMode {
   kBandPass,
 };
 
-#if CUDNN_VERSION < 6000
-#pragma message "CUDNN version under 6.0 is supported at best effort."
-#pragma message "We strongly encourage you to move to 6.0 and above."
-#pragma message "This message is intended to annoy you enough to update."
-#pragma message \
-    "please see https://docs.nvidia.com/deeplearning/sdk/cudnn-release-notes/"
-
-inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) {
-  switch (mode) {
-    case PoolingMode::kMaximumDeterministic:
-      return CUDNN_POOLING_MAX;
-    case PoolingMode::kAverageExclusive:
-      return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
-    case PoolingMode::kAverageInclusive:
-      return CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
-    case PoolingMode::kMaximum:
-      return CUDNN_POOLING_MAX;
-    default:
-      PADDLE_THROW(
-          platform::errors::Unimplemented("Unexpected CUDNN pooling mode."));
-  }
-}
-#else
-
 inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) {
   switch (mode) {
     case PoolingMode::kMaximumDeterministic:
@@ -130,7 +106,6 @@ inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) {
           platform::errors::Unimplemented("Unexpected CUDNN pooling mode."));
   }
 }
-#endif  // CUDNN_VERSION < 6000
 
 inline ActivationMode StringToActivationMode(const std::string& str) {
   if (str == "identity") {
@@ -471,19 +446,6 @@ class ScopedConvolutionDescriptor {
             "of pads is %d, size of dilations is %d.",
             pads.size(), dilations.size()));
 
-#if !CUDNN_VERSION_MIN(6, 0, 0)
-    // cudnn v5 does not support dilation conv, the argument is called upscale
-    // instead of dilations and it is must be one.
-    for (size_t i = 0; i < dilations.size(); ++i) {
-      PADDLE_ENFORCE_EQ(dilations[i], 1,
-                        platform::errors::InvalidArgument(
-                            "Dilations conv is not supported in this cuDNN "
-                            "version(%d.%d.%d).",
-                            CUDNN_VERSION / 1000, CUDNN_VERSION % 1000 / 100,
-                            CUDNN_VERSION % 100));
-    }
-#endif
-
     cudnnDataType_t compute_type =
         (type == CUDNN_DATA_DOUBLE) ? CUDNN_DATA_DOUBLE : CUDNN_DATA_FLOAT;
     PADDLE_ENFORCE_CUDA_SUCCESS(dynload::cudnnSetConvolutionNdDescriptor(