diff --git a/paddle/fluid/inference/analysis/passes/CMakeLists.txt b/paddle/fluid/inference/analysis/passes/CMakeLists.txt
index 9d74dc6c211e4fcb6d1e7de5369eee847f49fc78..a8d0c69a54ab39781613d26474098450398d4c1b 100644
--- a/paddle/fluid/inference/analysis/passes/CMakeLists.txt
+++ b/paddle/fluid/inference/analysis/passes/CMakeLists.txt
@@ -3,11 +3,13 @@ cc_library(ir_analysis_pass SRCS ir_analysis_pass.cc DEPS analysis_pass argument
 cc_library(memory_optim_pass SRCS memory_optimize_pass.cc DEPS analysis_pass zero_copy_tensor)
 cc_library(ir_params_sync_among_devices_pass SRCS ir_params_sync_among_devices_pass.cc DEPS analysis_pass argument ir_pass_manager)
 cc_library(ir_graph_to_program_pass SRCS ir_graph_to_program_pass.cc DEPS analysis_pass graph_to_program_pass)
+cc_library(adjust_cudnn_workspace_size_pass SRCS adjust_cudnn_workspace_size_pass.cc DEPS analysis_pass graph_to_program_pass)
 
 cc_library(analysis_passes SRCS passes.cc DEPS
   ir_graph_build_pass
   ir_analysis_pass
   ir_params_sync_among_devices_pass
+  adjust_cudnn_workspace_size_pass
   memory_optim_pass
   ir_graph_to_program_pass
 )
diff --git a/paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.cc b/paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.cc
new file mode 100644
index 0000000000000000000000000000000000000000..0470e0d5a247163ecd7e7dd1e8f88e6b71ae93d7
--- /dev/null
+++ b/paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.cc
@@ -0,0 +1,43 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+void AdjustCudnnWorkSpacePass::RunImpl(Argument* argument) {
+  if (!argument->use_gpu()) return;
+  auto& graph = argument->main_graph();
+  auto nodes = graph.Nodes();
+  const int cudnn_workspace_size_MB = 64;
+  const std::string attr_name = "workspace_size_MB";
+
+  for (auto& node : nodes) {
+    if (!node->IsOp()) continue;
+    auto* op_desc = node->Op();
+    if (!op_desc->HasAttr(attr_name)) continue;
+    op_desc->SetAttr(attr_name, cudnn_workspace_size_MB);
+    op_desc->Flush();
+  }
+}
+
+std::string AdjustCudnnWorkSpacePass::repr() const {
+  return "adjust-cudnn-work-space-pass";
+}
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle
diff --git a/paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.h b/paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.h
new file mode 100644
index 0000000000000000000000000000000000000000..65d1c545313e110028a92776e73a070d32010420
--- /dev/null
+++ b/paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.h
@@ -0,0 +1,41 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/inference/analysis/analysis_pass.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+/*
+ * The default cudnn workspace is 4G, we set it to 64M in this pass, which
+ * is applicable for most inference tasks.
+ */
+class AdjustCudnnWorkSpacePass : public AnalysisPass {
+ public:
+  void RunImpl(Argument *argument) override;
+  std::string repr() const override;
+};
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle
diff --git a/paddle/fluid/inference/analysis/passes/passes.cc b/paddle/fluid/inference/analysis/passes/passes.cc
index 161b127d6d5ceb3e8b9c1cf98c69eb0387bfb905..a55904ed536bad31c82888ede2db3178f3fd5e47 100644
--- a/paddle/fluid/inference/analysis/passes/passes.cc
+++ b/paddle/fluid/inference/analysis/passes/passes.cc
@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/inference/analysis/passes/passes.h"
+#include "paddle/fluid/inference/analysis/passes/adjust_cudnn_workspace_size_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.h"
@@ -35,6 +36,8 @@ PassRegistry::PassRegistry() {
   passes_.emplace(
       "ir_params_sync_among_devices_pass",
       std::unique_ptr<AnalysisPass>(new IrParamsSyncAmongDevicesPass));
+  passes_.emplace("adjust_cudnn_workspace_size_pass",
+                  std::unique_ptr<AnalysisPass>(new AdjustCudnnWorkSpacePass));
   passes_.emplace(
       "ir_graph_to_program_pass",
       std::unique_ptr<IrGraphToProgramPass>(new IrGraphToProgramPass));
diff --git a/paddle/fluid/inference/analysis/passes/passes.h b/paddle/fluid/inference/analysis/passes/passes.h
index ea07e0dcbd992c9d10c6662909798ef79a01e3a7..8a13091d083e51ecc84e6790f973ffa39ba5a6b9 100644
--- a/paddle/fluid/inference/analysis/passes/passes.h
+++ b/paddle/fluid/inference/analysis/passes/passes.h
@@ -14,7 +14,9 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include "paddle/fluid/inference/analysis/analysis_pass.h"
 
 namespace paddle {
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index 321107377c286eb4212b62be481a3417ef3f72d2..e57d3a80456767848143412b2524f94fa09c7c13 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -120,7 +120,11 @@ bool AnalysisPredictor::PrepareScope(
     scope_ = parent_scope;
     status_is_cloned_ = true;
   } else {
-    paddle::framework::InitDevices(false);
+    if (config_.use_gpu_) {
+      paddle::framework::InitDevices(false, {config_.device_id_});
+    } else {
+      paddle::framework::InitDevices(false, {});
+    }
     scope_.reset(new paddle::framework::Scope());
     status_is_cloned_ = false;
   }
@@ -459,6 +463,8 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
       std::string flag = "--fraction_of_gpu_memory_to_use=" +
                          std::to_string(fraction_of_gpu_memory);
       flags.push_back(flag);
+      flags.push_back("--selected_gpus=" +
+                      std::to_string(config.gpu_device_id()));
       VLOG(3) << "set flag: " << flag;
       framework::InitGflags(flags);
     }
diff --git a/paddle/fluid/inference/api/paddle_pass_builder.h b/paddle/fluid/inference/api/paddle_pass_builder.h
index 09ef195d5e66aff0cef17f1594de34c656187a35..057e7dc65d5fd41212cbee77a2a4f4431b011182 100644
--- a/paddle/fluid/inference/api/paddle_pass_builder.h
+++ b/paddle/fluid/inference/api/paddle_pass_builder.h
@@ -73,7 +73,8 @@ class PaddlePassBuilder {
  protected:
   std::vector<std::string> analysis_passes_{
       {"ir_graph_build_pass", "ir_analysis_pass",
-       "ir_params_sync_among_devices_pass"}};
+       "ir_params_sync_among_devices_pass",
+       "adjust_cudnn_workspace_size_pass"}};
   std::vector<std::string> passes_;
 };