Multiunit recording with multi-site electrodes in the brain has been widely used in neuroscience studies. After the data recording, neuronal spikes should be sorted according to the pattern of spike waveforms. For the spike sorting, independent component analysis (ICA) has recently been used because ICA has potential for resolving the problem to separate the overlapped multiple neuronal spikes. However the performance of spike sorting by using ICA has not been examined in detail. In this study, we quantitatively evaluate the performance of ICA-based spike sorting method by using simulated multiunit signals. The simulated multiunit signal is constructed by compositing real extracellular action potentials recorded from guinea-pig brain. It is found that the spike sorting by using ICA hardly avoids significant false positive and negative errors due to the cross-talk noise contamination on the separated signals. The cross-talk occurs when the multiunit signal of each recording channel have significant time difference; this situation does not satisfy the assumption of instantaneous source mixture for the major ICA algorithms. Since the channel delay problem is hardly resolved, an ICA algorithm which does not require the instantaneous source mixing assumption would be appropriate for use of spike sorting.