diff --git a/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.C b/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.C
index e305fbb0e..f7f4dced6 100644
--- a/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.C
+++ b/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.C
@@ -765,6 +765,13 @@ void colvarproxy_impl::calculate() {
   cudaCheck(cudaSetDevice(m_device_id));
   // The following memcpy operations are supposed to be overlapped with the NB kernel
   if (numAtoms > 0) {
+    if (mClient->requestUpdateAtomTotalForces()) {
+      if (mClient->isStartupStep()) {
+        set_total_forces_invalid();
+      } else {
+        set_total_forces_valid();
+      }
+    }
     if (!has_gpu_support()) {
       // Transform the arrays for Colvars
       auto &colvars_pos = *(modify_atom_positions());
@@ -1078,8 +1085,8 @@ bool CudaGlobalMasterColvars::requestUpdateCharges() {
   return mImpl->atomsChanged();
 }
 
-void CudaGlobalMasterColvars::setStep(int64_t step) {
-  CudaGlobalMasterClient::setStep(step);
+void CudaGlobalMasterColvars::setStep(int64_t step, int startup, int doMigration) {
+  CudaGlobalMasterClient::setStep(step, startup, doMigration);
   if (mImpl->atomsChanged()) {
     mImpl->reallocate();
   }
diff --git a/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.h b/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.h
index aa2042438..7000afa2c 100644
--- a/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.h
+++ b/namd/cudaglobalmaster/colvarproxy_cudaglobalmaster.h
@@ -34,7 +34,7 @@ class CudaGlobalMasterColvars: public CudaGlobalMasterClient {
   int updateFromTCLCommand(const std::vector<std::string>& arguments) override;
   void onBuffersUpdated() override;
   void calculate() override;
-  void setStep(int64_t step) override;
+  void setStep(int64_t step, int startup, int doMigration) override;
   cudaStream_t getStream() override;
   bool requestedAtomsChanged() override;
   bool requestedForcedAtomsChanged() override { return requestedAtomsChanged(); }
@@ -76,6 +76,7 @@ class CudaGlobalMasterColvars: public CudaGlobalMasterClient {
     return CudaGlobalMasterClient::replica_comm_send(msg_data, msg_len, dest_rep);
   }
   int64_t getStep() const {return m_step;}
+  int isStartupStep() const {return m_is_startup_step;}
   std::string getTCLUpdateResult() override {return mTCLResult;}
 private:
   std::unique_ptr<colvarproxy_impl> mImpl;
diff --git a/src/colvar_gpu_calc.cpp b/src/colvar_gpu_calc.cpp
index c0a9bd82c..93328317a 100644
--- a/src/colvar_gpu_calc.cpp
+++ b/src/colvar_gpu_calc.cpp
@@ -708,9 +708,7 @@ int colvarmodule_gpu_calc::calc_cvs(const std::vector<colvar*>& colvars, colvarm
   // Update flags
   checkColvarsError(cv_update_flags(colvars));
   // Calculate total force
-  if (colvar_module->step_relative() > 0) {
-    checkColvarsError(cvc_calc_total_force(colvars, calc_total_force_compute, colvar_module, false));
-  }
+  checkColvarsError(cvc_calc_total_force(colvars, calc_total_force_compute, colvar_module, false));
   // Read data to atom groups
   checkColvarsError(atom_group_read_data_gpu(
     colvars, read_data_compute, colvar_module));