diff --git a/mace/core/mace.cc b/mace/core/mace.cc
index 2444f66365682ac89567a0116250e759c280a7da..3e64e2b1a60c7c423d8e063f8f70caeda2232b77 100644
--- a/mace/core/mace.cc
+++ b/mace/core/mace.cc
@@ -355,7 +355,7 @@ void ConfigOpenCLRuntime(GPUPerfHint gpu_perf_hint,
                          GPUPriorityHint gpu_priority_hint) {
   VLOG(1) << "Set GPU configurations, gpu_perf_hint: " << gpu_perf_hint
           << ", gpu_priority_hint: " << gpu_priority_hint;
-  OpenCLRuntime::CreateGlobal(gpu_perf_hint, gpu_priority_hint);
+  OpenCLRuntime::Configure(gpu_perf_hint, gpu_priority_hint);
 }
 
 void ConfigOmpThreadsAndAffinity(int omp_num_threads,
diff --git a/mace/core/runtime/opencl/opencl_runtime.cc b/mace/core/runtime/opencl/opencl_runtime.cc
index c5b6557bd550d9c2d9f88a130f76aeb7e725ff22..798409abf683c489104a96a80bf9938b906e3d4f 100644
--- a/mace/core/runtime/opencl/opencl_runtime.cc
+++ b/mace/core/runtime/opencl/opencl_runtime.cc
@@ -65,23 +65,19 @@ void OpenCLProfilingTimer::ClearTiming() {
   accumulated_micros_ = 0;
 }
 
-std::unique_ptr<OpenCLRuntime> OpenCLRuntime::runtime_instance_ = nullptr;
+GPUPerfHint OpenCLRuntime::gpu_perf_hint_ = GPUPerfHint::PERF_DEFAULT;
+GPUPriorityHint OpenCLRuntime::gpu_priority_hint_ =
+    GPUPriorityHint::PRIORITY_DEFAULT;
 
 OpenCLRuntime *OpenCLRuntime::Global() {
-  // FIXME: not thread safe
-  if (runtime_instance_ == nullptr) {
-    return CreateGlobal(GPUPerfHint::PERF_DEFAULT,
-                        GPUPriorityHint::PRIORITY_DEFAULT);
-  }
-  return runtime_instance_.get();
+  static OpenCLRuntime runtime(gpu_perf_hint_, gpu_priority_hint_);
+  return &runtime;
 }
 
-OpenCLRuntime *OpenCLRuntime::CreateGlobal(GPUPerfHint gpu_perf_hint,
-                                           GPUPriorityHint gpu_priority_hint) {
-  runtime_instance_ =
-      std::unique_ptr<OpenCLRuntime>(new OpenCLRuntime(gpu_perf_hint,
-                                                       gpu_priority_hint));
-  return runtime_instance_.get();
+void OpenCLRuntime::Configure(GPUPerfHint gpu_perf_hint,
+                              GPUPriorityHint gpu_priority_hint) {
+  OpenCLRuntime::gpu_perf_hint_ = gpu_perf_hint;
+  OpenCLRuntime::gpu_priority_hint_ = gpu_priority_hint;
 }
 
 void GetAdrenoContextProperties(std::vector<cl_context_properties> *properties,
diff --git a/mace/core/runtime/opencl/opencl_runtime.h b/mace/core/runtime/opencl/opencl_runtime.h
index 801ee4711dcf07441ece9dcca4bc3ce13c3ff65f..ce375b9aed7bafe47ce7cd97310a76e496008544 100644
--- a/mace/core/runtime/opencl/opencl_runtime.h
+++ b/mace/core/runtime/opencl/opencl_runtime.h
@@ -39,7 +39,7 @@ class OpenCLProfilingTimer : public Timer {
 class OpenCLRuntime {
  public:
   static OpenCLRuntime *Global();
-  static OpenCLRuntime *CreateGlobal(GPUPerfHint, GPUPriorityHint);
+  static void Configure(GPUPerfHint, GPUPriorityHint);
 
   cl::Context &context();
   cl::Device &device();
@@ -52,10 +52,10 @@ class OpenCLRuntime {
   cl::Kernel BuildKernel(const std::string &program_name,
                          const std::string &kernel_name,
                          const std::set<std::string> &build_options);
-  ~OpenCLRuntime();
 
  private:
   OpenCLRuntime(GPUPerfHint, GPUPriorityHint);
+  ~OpenCLRuntime();
   OpenCLRuntime(const OpenCLRuntime &) = delete;
   OpenCLRuntime &operator=(const OpenCLRuntime &) = delete;
 
@@ -74,7 +74,9 @@ class OpenCLRuntime {
   std::map<std::string, cl::Program> built_program_map_;
   std::mutex program_build_mutex_;
   std::string kernel_path_;
-  static std::unique_ptr<OpenCLRuntime> runtime_instance_;
+
+  static GPUPerfHint gpu_perf_hint_;
+  static GPUPriorityHint gpu_priority_hint_;
 };
 
 }  // namespace mace