diff --git a/examples/cfd/flow_past_sphere_3d.py b/examples/cfd/flow_past_sphere_3d.py
index b74128ae..2872d7fe 100644
--- a/examples/cfd/flow_past_sphere_3d.py
+++ b/examples/cfd/flow_past_sphere_3d.py
@@ -158,6 +158,8 @@ def post_process(step, f_current):
     f_0, f_1 = f_1, f_0  # Swap the buffers
 
     if step % post_process_interval == 0 or step == num_steps - 1:
+        if compute_backend == ComputeBackend.WARP:
+            wp.synchronize()
         post_process(step, f_0)
         end_time = time.time()
         elapsed = end_time - start_time
diff --git a/examples/cfd/turbulent_channel_3d.py b/examples/cfd/turbulent_channel_3d.py
index 599a9f26..0abfbef0 100644
--- a/examples/cfd/turbulent_channel_3d.py
+++ b/examples/cfd/turbulent_channel_3d.py
@@ -206,9 +206,11 @@ def plot_uplus(u, timestep, grid_shape, u_tau, visc):
     f_0, f_1 = stepper(f_0, f_1, bc_mask, missing_mask, omega, step)
     f_0, f_1 = f_1, f_0  # Swap the buffers
 
-    if (step + 1) % print_interval == 0:
+    if step % print_interval == 0:
+        if compute_backend == ComputeBackend.WARP:
+            wp.synchronize()
         elapsed_time = time.time() - start_time
-        print(f"Iteration: {step + 1}/{num_steps} | Time elapsed: {elapsed_time:.2f}s")
+        print(f"Iteration: {step}/{num_steps} | Time elapsed: {elapsed_time:.2f}s")
         start_time = time.time()
 
     if (step % post_process_interval == 0) or (step == num_steps - 1):
diff --git a/examples/cfd/windtunnel_3d.py b/examples/cfd/windtunnel_3d.py
index 1f0d266e..074570b8 100644
--- a/examples/cfd/windtunnel_3d.py
+++ b/examples/cfd/windtunnel_3d.py
@@ -199,12 +199,12 @@ def post_process(
 
     # Compute lift and drag
     boundary_force = momentum_transfer(f_0, f_1, bc_mask, missing_mask)
-    drag = np.sqrt(boundary_force[0] ** 2 + boundary_force[1] ** 2)  # xy-plane
+    drag = boundary_force[0]  # x-direction
     lift = boundary_force[2]
-    c_d = 2.0 * drag / (wind_speed**2 * car_cross_section)
-    c_l = 2.0 * lift / (wind_speed**2 * car_cross_section)
-    drag_coefficients.append(c_d)
-    lift_coefficients.append(c_l)
+    cd = 2.0 * drag / (wind_speed**2 * car_cross_section)
+    cl = 2.0 * lift / (wind_speed**2 * car_cross_section)
+    drag_coefficients.append(cd)
+    lift_coefficients.append(cl)
     time_steps.append(step)
 
     # Plot drag coefficient
@@ -236,9 +236,11 @@ def post_process(
     f_0, f_1 = f_1, f_0  # Swap the buffers
 
     # Print progress at intervals
-    if (step + 1) % print_interval == 0:
+    if step % print_interval == 0:
+        if compute_backend == ComputeBackend.WARP:
+            wp.synchronize()
         elapsed_time = time.time() - start_time
-        print(f"Iteration: {step + 1}/{num_steps} | Time elapsed: {elapsed_time:.2f}s")
+        print(f"Iteration: {step}/{num_steps} | Time elapsed: {elapsed_time:.2f}s")
         start_time = time.time()
 
     # Post-process at intervals and final step