Definition
Plain language
One run of input through a neural network from start to finish to produce an output.
As stated in the literature
A single application of a network's parameters to an input, computing activations layer-by-layer to produce predictions.
Why it matters: It's the basic unit of inference cost, and most performance and latency discussions ultimately reduce to how many forward passes you're doing.
For example, when you type a question to a chatbot, the model does one forward pass per output token to decide what word comes next.
Heard on the show
“Judging is one decision — one forward pass, one bit out.”Episode 233 — Why a Model Can Grade an Answer But Not Write the Answer Key