diff --git a/cuda_bindings/cuda/bindings/_lib/utils.pxi b/cuda_bindings/cuda/bindings/_lib/utils.pxi
index 7783afed97c..7e4b1185fc3 100644
--- a/cuda_bindings/cuda/bindings/_lib/utils.pxi
+++ b/cuda_bindings/cuda/bindings/_lib/utils.pxi
@@ -308,11 +308,12 @@ cdef class _HelperCUpointer_attribute:
         if self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_CONTEXT,):
             return self._ctx
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_MEMORY_TYPE,
-                            cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ACCESS_FLAGS,):
             return self._uint
+        elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,):
+            return self._int
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_POINTER,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_RANGE_START_ADDR,):
             return self._devptr
diff --git a/cuda_bindings/tests/test_cuda.py b/cuda_bindings/tests/test_cuda.py
index 7bef2b844aa..bd871bcd5eb 100644
--- a/cuda_bindings/tests/test_cuda.py
+++ b/cuda_bindings/tests/test_cuda.py
@@ -1335,3 +1335,35 @@ def test_dealloc_clears_array_field_in_external_struct():
         f"external struct still holds a dangling pointer ({attrs_after:#x}) "
         "where attrs was, after the aliasing wrapper was destroyed"
     )
+
+
+def test_pointer_device_ordinal_matches_allocation_device():
+    # Regression test for cuPointerGetAttribute(s)(DEVICE_ORDINAL): the binding
+    # must report the real device the pointer was allocated on, not always 0.
+    # Needs >= 2 GPUs: on device 0 a wrong 0 is indistinguishable from correct.
+    err, ndev = cudart.cudaGetDeviceCount()
+    assert err == cudart.cudaError_t.cudaSuccess
+    if ndev < 2:
+        pytest.skip("needs >= 2 GPUs to distinguish a wrong 0 from the correct ordinal")
+
+    ord_attr = cuda.CUpointer_attribute.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL
+    for dev in range(ndev):
+        (err,) = cudart.cudaSetDevice(dev)
+        assert err == cudart.cudaError_t.cudaSuccess
+        err, ptr = cudart.cudaMalloc(256)
+        assert err == cudart.cudaError_t.cudaSuccess
+        try:
+            err, attrs = cudart.cudaPointerGetAttributes(ptr)
+            assert err == cudart.cudaError_t.cudaSuccess
+            expected = attrs.device
+
+            err, singular = cuda.cuPointerGetAttribute(ord_attr, cuda.CUdeviceptr(ptr))
+            assert err == cuda.CUresult.CUDA_SUCCESS
+            assert singular == expected, f"singular ordinal {singular} != {expected} on device {dev}"
+
+            err, plural = cuda.cuPointerGetAttributes(1, [ord_attr], cuda.CUdeviceptr(ptr))
+            assert err == cuda.CUresult.CUDA_SUCCESS
+            assert plural[0] == expected, f"plural ordinal {plural[0]} != {expected} on device {dev}"
+        finally:
+            (err,) = cudart.cudaFree(ptr)
+            assert err == cudart.cudaError_t.cudaSuccess