[andrew-wja]

Primarily it's down to the fact that most programs have control flow!

As you most likely know, DAG selectors have to tile the CFG (which can have back-edges due to control flow) into a forest of DAGs, which are then fed to the instruction selector. So from a big program, you get a bunch of DAGs. The DAGs are selected separately, which means that when they execute one-after-the-other or in parallel, there may be inter-DAG conflicts on things like issue ports, or memory ports in ALUs. I'm sure there are a billion mitigation strategies for this, but for DNNs the landscape is a little different.

With a DNN you may have parallel paths, but they are typically few, and there are no back-edges. So the entire program is a DAG, and you don't have to tile and select separately!