Mammalian rods respond to single photons with a hyperpolarization of about 1 mV which is accompanied by continuous noise. Since the mammalian rod bipolar cell collects signals from 20–100 rods, the noise from the converging rods would overwhelm the single-photon signal from one rod at scotopic intensities (starlight) if the bipolar cell summed signals linearly (Baylor et al., 1984). However, it is known that at scotopic intensities the retina preserves single-photon responses (Barlow et al., 1971; Mastronarde, 1983). To explore noise summation in the rod bipolar pathway, we simulated an array of rods synaptically connected to a rod bipolar cell using a compartmental model. The performance of the circuit was evaluated with a discriminator measuring errors in photon detection as false positives and false negatives, which were compared to physiologically and psychophysically measured error rates. When only one rod was connected to the rod bipolar, a Poisson rate of 80 vesicles/s was necessary for reliable transmission of the single-photon signal. When 25 rods converged through a linear synapse the noise caused an unacceptably high false positive rate, even when either dark continuous noise or synaptic noise where completely removed. We propose that a threshold nonlinearity is provided by the mGluR6 receptor in the rod bipolar dendrite (Shiells & Falk, 1994) to yield a synapse with a noise removing mechanism. With the threshold nonlinearity the synapse removed most of the noise. These results suggest that a threshold provided by the mGluR6 receptor in the rod bipolar cell is necessary for proper functioning of the retina at scotopic intensities and that the metabotropic domains in the rod bipolar are distinct. Such a nonlinear threshold could also reduce synaptic noise for cortical circuits in which sparse signals converge.