How I understand the tangle is that each transaction is a unique vertex in the directed acyclic graph. Instead of a block that contains multiple transactions with a corresponding hash of the transactions in a block, you have a path to a source that gives the initial creation of the token, i.e. from an ICO or from mining (if that is done with the tangle you're using).
A tip (or equivalently a sink in the DAG) is confirmed based on the consensus model (some type of max sum that aggregates all the weight-based paths).
I would think that you can process this in parallel efficiently, by stopping at each vertex that has multiple transaction outputs and waiting for each path to merge back to it, by first starting at the tips and tracing the paths backwards to the source.
Granted, I won't claim to know anything about tangles other than fascination and a desire to learn and understand alternative consensus models.
But, this is my vague high-level overview of what's happening.