The retina's visual message is transmitted to the brain by ganglion cells that integrate noisy synaptic inputs to create a spike train. We asked how efficiently the retinal ganglion cell spike generator creates the spike train message. Intracellular and extracellular recording were taken from in vitro guinea pig retina, in response to a spot of light flashed over the receptive field center. Responses were analyzed with an "ideal observer," a program that discriminated between two contrasts based on an optimal decisin rule. Spike trains from ganglion cells had detection thresholds as low as 1% contrast, but thresholds for the corresponding graded potentials were lower by a factor of 2. Using a computational model of the ganglion cell,we asked what factors in the spike generator mechanism are responsible for the spike train's loss in performance. The model included dendritic/axonal morphology, noisy synaptic inputs and 7 types of membrane channels. Adaptation of spike rate was provided by K(Ca) channels which were activated by Ca²⁺ flux during spikes. When K(Ca) channels were included, they controlled the duration of the inter-spike interval and thus set the level of noise in the spike train. These results imply that the spike generator adds noise to the spike train signal.