diff --git a/lidar_pipeline/visualizations.py b/lidar_pipeline/visualizations.py
index a06dcd3..041dffd 100644
--- a/lidar_pipeline/visualizations.py
+++ b/lidar_pipeline/visualizations.py
@@ -298,6 +298,11 @@ def _ray_trace_horizons(dem, rows, cols, res, n_dirs, max_dist, radii_m=None):
     recording the max upward angle (positive openness) and max downward angle
     (negative openness) reached at each radius checkpoint.
 
+    Padding is done on CPU (numpy) to avoid GPU memory pressure and
+    pre-compiled kernel mismatches (CUDA_ERROR_NO_BINARY_FOR_GPU on sm_89).
+    The padded array is transferred to GPU once, then each direction is
+    processed and results are streamed back to CPU.
+
     Args:
         dem: GPU or CPU filled DEM array (rows, cols).
         rows, cols: dimensions.
@@ -322,7 +327,15 @@ def _ray_trace_horizons(dem, rows, cols, res, n_dirs, max_dist, radii_m=None):
         radii_steps = [max_dist]
         n_radii = 1
 
-    padded = xp.pad(dem, max_dist, mode='constant', constant_values=xp.nan)
+    # Pad on CPU (numpy) — avoids GPU memory pressure and
+    # pre-compiled kernel issues (NO_BINARY_FOR_GPU on sm_89).
+    dem_np = to_cpu(dem)
+    padded_np = np.pad(dem_np, max_dist, mode='constant', constant_values=np.nan)
+
+    # Transfer padded DEM to GPU for computation
+    padded = to_gpu(padded_np)
+    # Free the CPU copy — we don't need it anymore
+    del dem_np, padded_np
 
     # Process one direction at a time to limit GPU memory.
     # Store results as flat CPU arrays — transfer back to GPU at the end.
@@ -349,13 +362,17 @@ def _ray_trace_horizons(dem, rows, cols, res, n_dirs, max_dist, radii_m=None):
         radii_remaining = set(range(n_radii))
 
         for step, px, py, dist_m in valid_steps:
-            elev_diff = padded[max_dist + py:max_dist + py + rows,
-                              max_dist + px:max_dist + px + cols] - dem
+            # Slice from padded array, subtract original dem
+            view = padded[max_dist + py:max_dist + py + rows,
+                         max_dist + px:max_dist + px + cols]
+            elev_diff = view - dem
+            del view  # free slice reference
 
             # Positive: angle to terrain above viewer
             pos_angle = xp.arctan2(xp.maximum(elev_diff, 0), dist_m)
             # Negative: angle to terrain below viewer
             neg_angle = xp.arctan2(xp.maximum(-elev_diff, 0), dist_m)
+            del elev_diff  # free intermediate
 
             # Update running max for all radius checkpoints still active
             for r_idx in radii_remaining: