The stated propagation delay for TTL chips was usually about 10 or 20 nanoseconds (sometimes larger for more complex logical operations, and frequently smaller in subsequent logic generations, like 5 nanoseconds or so). I just checked this in a late 1970s databook.
So, 20 MHz seems pretty credible for a region where you would see problems due to propagation delay—maybe even somewhere below that with the original 7400 series.
That's also a reason that I hadn't thought about so much why putting the whole CPU on a single die would look more and more practical over time: propagation delay in wires!