diff --git a/src/include/npkit/npkit.h b/src/include/npkit/npkit.h
index 4e208050..a0691afd 100644
--- a/src/include/npkit/npkit.h
+++ b/src/include/npkit/npkit.h
@@ -40,11 +40,9 @@ public:
 
   static void CollectCpuEvent(uint8_t type, uint32_t size, uint32_t rsvd, uint64_t timestamp, int channel_id);
 
-  static uint64_t* GetCpuTimestamp();
+  static uint64_t GetCpuTimestamp();
 
 private:
-  static void CpuTimestampUpdateThread();
-
   // 64K * 512 * 16B = 512MB per GPU
   static const uint64_t kMaxNumGpuEventsPerBuffer = 1ULL << 16;
 
@@ -56,12 +54,11 @@ private:
 
   static NpKitEventCollectContext* gpu_collect_contexts_;
   static NpKitEventCollectContext* cpu_collect_contexts_;
-  static uint64_t* cpu_timestamp_;
+
+  static uint64_t cpu_base_system_timestamp_;
+  static uint64_t cpu_base_steady_timestamp_;
 
   static uint64_t rank_;
-
-  static std::thread* cpu_timestamp_update_thread_;
-  static volatile bool cpu_timestamp_update_thread_should_stop_;
 };
 
 #endif
\ No newline at end of file
diff --git a/src/misc/npkit.cc b/src/misc/npkit.cc
index 99a8cde8..4a7eb849 100644
--- a/src/misc/npkit.cc
+++ b/src/misc/npkit.cc
@@ -12,22 +12,8 @@ NpKitEvent** NpKit::cpu_event_buffers_ = nullptr;
 
 NpKitEventCollectContext* NpKit::gpu_collect_contexts_ = nullptr;
 NpKitEventCollectContext* NpKit::cpu_collect_contexts_ = nullptr;
-uint64_t* NpKit::cpu_timestamp_ = nullptr;
-
-std::thread* NpKit::cpu_timestamp_update_thread_ = nullptr;
-volatile bool NpKit::cpu_timestamp_update_thread_should_stop_ = false;
-
-void NpKit::CpuTimestampUpdateThread()
-{
-  uint64_t init_system_clock = std::chrono::system_clock::now().time_since_epoch().count();
-  uint64_t init_steady_clock = std::chrono::steady_clock::now().time_since_epoch().count();
-  uint64_t curr_steady_clock = 0;
-  volatile uint64_t* volatile_cpu_timestamp_ = cpu_timestamp_;
-  while (!cpu_timestamp_update_thread_should_stop_) {
-    curr_steady_clock = std::chrono::steady_clock::now().time_since_epoch().count();
-    *volatile_cpu_timestamp_ = init_system_clock + (curr_steady_clock - init_steady_clock);
-  }
-}
+uint64_t NpKit::cpu_base_system_timestamp_ = 0;
+uint64_t NpKit::cpu_base_steady_timestamp_ = 0;
 
 mscclppResult_t NpKit::Init(int rank)
 {
@@ -54,11 +40,8 @@ mscclppResult_t NpKit::Init(int rank)
   }
 
   // Init timestamp
-  MSCCLPPCHECK(mscclppCudaHostCalloc(&cpu_timestamp_, 1));
-  volatile uint64_t* volatile_cpu_timestamp = cpu_timestamp_;
-  *volatile_cpu_timestamp = std::chrono::system_clock::now().time_since_epoch().count();
-  cpu_timestamp_update_thread_should_stop_ = false;
-  cpu_timestamp_update_thread_ = new std::thread(CpuTimestampUpdateThread);
+  cpu_base_system_timestamp_ = std::chrono::system_clock::now().time_since_epoch().count();
+  cpu_base_steady_timestamp_ = std::chrono::steady_clock::now().time_since_epoch().count();
 
   return mscclppSuccess;
 }
@@ -133,10 +116,6 @@ mscclppResult_t NpKit::Shutdown()
 {
   uint64_t i = 0;
 
-  // Stop CPU timestamp updating thread
-  cpu_timestamp_update_thread_should_stop_ = true;
-  cpu_timestamp_update_thread_->join();
-
   // Free CPU event data structures
   for (i = 0; i < kNumCpuEventBuffers; i++) {
     free(cpu_event_buffers_[i]);
@@ -151,9 +130,6 @@ mscclppResult_t NpKit::Shutdown()
   free(gpu_event_buffers_);
   CUDACHECK(cudaFree(gpu_collect_contexts_));
 
-  // Free timestamp
-  MSCCLPPCHECK(mscclppCudaHostFree(cpu_timestamp_));
-
   return mscclppSuccess;
 }
 
@@ -175,7 +151,8 @@ void NpKit::CollectCpuEvent(uint8_t type, uint32_t size, uint32_t rsvd, uint64_t
   }
 }
 
-uint64_t* NpKit::GetCpuTimestamp()
+uint64_t NpKit::GetCpuTimestamp()
 {
-  return cpu_timestamp_;
-}
\ No newline at end of file
+  uint64_t cpu_curr_steady_timestamp_ = std::chrono::steady_clock::now().time_since_epoch().count();
+  return cpu_base_steady_timestamp_ + (cpu_curr_steady_timestamp_ - cpu_base_steady_timestamp_);
+}
diff --git a/src/proxy.cc b/src/proxy.cc
index 8d2fe6be..2c670fb6 100644
--- a/src/proxy.cc
+++ b/src/proxy.cc
@@ -62,7 +62,6 @@ static void npkitInitReqIds(struct mscclppComm* comm)
 
 static void npkitCollectEntryEvent(struct mscclppConn* conn, uint8_t type, uint32_t size, int channelId)
 {
-  uint64_t ts = *(volatile uint64_t*)NpKit::GetCpuTimestamp();
   uint64_t reqId = 0;
   if (conn->npkitFreeReqIds.size() == 0) {
     reqId = conn->npkitUsedReqIds.size();
@@ -71,15 +70,14 @@ static void npkitCollectEntryEvent(struct mscclppConn* conn, uint8_t type, uint3
     conn->npkitFreeReqIds.pop_back();
   }
   conn->npkitUsedReqIds.push_back(reqId);
-  NpKit::CollectCpuEvent(type, size, (uint32_t)reqId, ts, channelId);
+  NpKit::CollectCpuEvent(type, size, (uint32_t)reqId, NpKit::GetCpuTimestamp(), channelId);
 }
 
 static void npkitCollectExitEvents(struct mscclppConn* conn, uint8_t type, int channelId)
 {
-  uint64_t ts = *(volatile uint64_t*)NpKit::GetCpuTimestamp();
   while (conn->npkitUsedReqIds.size()) {
     uint64_t reqId = conn->npkitUsedReqIds.back();
-    NpKit::CollectCpuEvent(type, 0, (uint32_t)reqId, ts, channelId);
+    NpKit::CollectCpuEvent(type, 0, (uint32_t)reqId, NpKit::GetCpuTimestamp(), channelId);
     conn->npkitFreeReqIds.push_back(reqId);
     conn->npkitUsedReqIds.pop_back();
   }