}Forward latency (FL) – block level
}Time between
tokens consumed at inputs
and generated outputs
}Inputs tokens spread apart
to avoid congestion slowing down results
}May be data-dependent
}(Forward) latency – system
}A sum of forward latencies through blocks
}Must account for causality
of output tokens within blocks
}Earlier/latest arriving
token may cause output token
}I.e., notion of critical path exists