> generally we would expect a parallel algorithm to be no more energy efficient (with respect to total conceptual work performed -- e.g., training a fixed neura

> generally we would expect a parallel algorithm to be no more energy efficient (with respect to total conceptual work performed -- e.g., training a fixed neural network) than an equivalent serial implementation I think it depends on how you conceptualize the topology in both cases. A serial implementation requires threading all the data through a single point, whereas a parallel implementation can leave data where it is going to be used. Moving data around requires energy, so implementations that maximize locality of data should be more energy efficient. Such implementations would naturally synchronize as little as possible, so they would be highly parallel. Basically, serial implementations of neural networks require a clock and a form of RAM, including the energy overhead of dispatch, synchronization and data transport, whereas parallel implementations don't: each neuron could just contain whatever little data it needs and nothing more.