Neural information processing is shaped and constrained by the biophysics of neural circuitry. In particular, information is metabolically expensive to process and transmit, suggesting that neural codes should be energy efficient. In this talk I develop the theory of energy-efficient information processing and describe an algorithm for deriving the distribution of symbols in a code that optimizes information transmitted per unit energy. Numerical analysis shows that noise in the information transmission channel has a dramatic effect on the optimal code. I apply this theory to information processing by ganglion cells in the retina. Various ganglion cell types are known to transmit information in discrete firing events or bursts, consisting of a set of stereotyped voltage spikes. I will regard these bursts as symbols employed by the ganglion cell code and measure noise in their transmission by recording responses to repeated presentations of the same stimuli. Using this measured noise, and assuming a cost for bursts that is linear in their size, the theory of energy efficient codes predicts a specific distribution of coding symbols that optimizes information per energy in the ganglion cell. We will see that there is excellent qualitative and quantitative agreement between this prediction and the observed behaviour of retinal ganglion cells.