diff --git a/lidar_pipeline/gpu.py b/lidar_pipeline/gpu.py
index ad9e72c..5b3650b 100644
--- a/lidar_pipeline/gpu.py
+++ b/lidar_pipeline/gpu.py
@@ -91,6 +91,14 @@ def _init_gpu():
         _xp = _real_cupy
         _cp = _real_cupy
         _cp_ndimage = _real_cupy_ndimage
+        # Limit GPU memory pool per worker to avoid OOM when multiple
+        # workers share one GPU.  Each worker gets at most 3.5 GB (or
+        # 50 % of total VRAM on smaller cards).
+        props = _real_cupy.cuda.runtime.getDeviceProperties(0)
+        total_mem = props['totalGlobalMem']
+        max_worker_mem = min(int(total_mem * 0.5), 3.5 * 1024**3)
+        _real_cupy.cuda.set_memory_pool(0, max_worker_mem)
+        logger.info(f"  Pool mémoire GPU limité à {max_worker_mem // (1024**3) * 1000 // 1024} MB")
     except (ImportError, Exception) as e:
         logger.warning(f"GPU non disponible — mode CPU: {e}")
         _xp = np