diff --git a/src/spikeinterface/preprocessing/common_reference.py b/src/spikeinterface/preprocessing/common_reference.py
index 6da823d0e4..79c939e759 100644
--- a/src/spikeinterface/preprocessing/common_reference.py
+++ b/src/spikeinterface/preprocessing/common_reference.py
@@ -247,11 +247,22 @@ def __init__(
         self.operator_func = np.mean if self.operator == "average" else np.median
 
     def get_traces(self, start_frame, end_frame, channel_indices):
+        # We need all the channels to calculate the reference
+        traces = self.parent_recording_segment.get_traces(start_frame, end_frame, slice(None))
+
+        # The reference of a sample only uses the channels at that sample, so the samples can be processed in
+        # blocks. This keeps the float copies made while re-referencing the size of a block instead of the chunk.
+        num_samples_per_block = 1024
+        num_channels = np.arange(traces.shape[1])[channel_indices].size
+        re_referenced_traces = np.empty((traces.shape[0], num_channels), dtype=self.dtype)
+        for block_start in range(0, traces.shape[0], num_samples_per_block):
+            block = slice(block_start, block_start + num_samples_per_block)
+            re_referenced_traces[block] = self._re_reference(traces[block], channel_indices)
+        return re_referenced_traces
+
+    def _re_reference(self, traces, channel_indices):
         # Let's do the case with group_indices equal None as that is easy
         if self.group_indices is None:
-            # We need all the channels to calculate the reference
-            traces = self.parent_recording_segment.get_traces(start_frame, end_frame, slice(None))
-
             if self.reference == "global":
                 if self.ref_channel_indices is None:
                     shift = self.operator_func(traces, axis=1, keepdims=True)
@@ -284,9 +295,6 @@ def get_traces(self, start_frame, end_frame, channel_indices):
 
         # Then the old implementation for backwards compatibility that supports grouping
         else:
-            # need input trace
-            traces = self.parent_recording_segment.get_traces(start_frame, end_frame, slice(None))
-
             sliced_channel_indices = np.arange(traces.shape[1])
             if channel_indices is not None:
                 sliced_channel_indices = sliced_channel_indices[channel_indices]