diff --git a/python/cudf/cudf/core/column/timedelta.py b/python/cudf/cudf/core/column/timedelta.py
index e6a466e522c..9333857fd09 100644
--- a/python/cudf/cudf/core/column/timedelta.py
+++ b/python/cudf/cudf/core/column/timedelta.py
@@ -4,6 +4,7 @@
 
 import datetime
 import functools
+import math
 from typing import TYPE_CHECKING, cast
 
 import numpy as np
@@ -39,20 +40,6 @@
     "D": 86_400_000_000_000,
 }
 
-_dtype_total_seconds_factor = {
-    np.dtype("timedelta64[s]"): 1.0,
-    np.dtype("timedelta64[ms]"): 1e-3,
-    np.dtype("timedelta64[us]"): 1e-6,
-    np.dtype("timedelta64[ns]"): 1e-9,
-}
-
-_dtype_total_seconds_decimal_round = {
-    np.dtype("timedelta64[s]"): 1,
-    np.dtype("timedelta64[ms]"): 3,
-    np.dtype("timedelta64[us]"): 6,
-    np.dtype("timedelta64[ns]"): 9,
-}
-
 
 class TimeDeltaColumn(ColumnBase):
     """
@@ -277,9 +264,15 @@ def time_unit(self) -> str:
         return np.datetime_data(self.dtype)[0]
 
     def total_seconds(self) -> ColumnBase:
+        conversion = _unit_to_nanoseconds_conversion[self.time_unit] / 1e9
+        # Typecast to decimal128 to avoid floating point precision issues
+        # https://github.com/rapidsai/cudf/issues/17664
         return (
-            self.astype("int64") * _dtype_total_seconds_factor[self.dtype]
-        ).round(decimals=_dtype_total_seconds_decimal_round[self.dtype])
+            (self.astype("int64") * conversion)
+            .astype(cudf.Decimal128Dtype(38, 9))
+            .round(decimals=abs(int(math.log10(conversion))))
+            .astype("float64")
+        )
 
     def ceil(self, freq: str) -> ColumnBase:
         raise NotImplementedError("ceil is currently not implemented")
diff --git a/python/cudf/cudf/tests/test_timedelta.py b/python/cudf/cudf/tests/test_timedelta.py
index 628d906e0b7..3e7e8a64911 100644
--- a/python/cudf/cudf/tests/test_timedelta.py
+++ b/python/cudf/cudf/tests/test_timedelta.py
@@ -1522,12 +1522,6 @@ def test_timedelta_series_total_seconds(data, dtype):
 @pytest.mark.parametrize("data", _TIMEDELTA_DATA)
 @pytest.mark.parametrize("dtype", utils.TIMEDELTA_TYPES)
 def test_timedelta_index_total_seconds(request, data, dtype):
-    request.applymarker(
-        pytest.mark.xfail(
-            condition=(1132.324 in data and dtype == "timedelta64[ms]"),
-            reason="https://github.com/rapidsai/cudf/issues/17664",
-        )
-    )
     gi = cudf.Index(data, dtype=dtype)
     pi = gi.to_pandas()