So if the paper is correct then you need about a half a million qubits?
Is this roughly a 40x improvement on the 20M qubits in 8 hours or is there more to it?
Shor's algorithm requires performing a modular exponentiation under superposition. For an n bit modulus this requires 2n or 3n qubits of storage, plus let's say 50% overhead for routing and gates. You end up needing 5n to 10n logical qubits for an n bit number. So to factor a 2048 bit number you'd need on the order of ten thousand logical qubits. Improving that to a few hundred logical qubits would be a big improvement. Also, there's fewer operations so the code distance can be lower.
...but don't forget that "if the paper is correct" bit.