That's 2.5 days on the 13 MW Summit supercomputer, for a single run. Presumably you would need multiple runs to get good statistics.

No, the classical simulation should give you the whole wavefunction, from which you can get the full probability distribution. One run should be enough.

Does it scale linearly? Is it 5 days on a half as good super computer? They just need to show that it can be done in any reasonable finite amount of time, not 2.5 days.

The main bottleneck is memory/storage capacity, not speed. Beyond that (if you have the requisite storage capacity and bandwidth), yes: if you have half the FLOPS and double the time, you get the same result.

Is this the case? I thought that parallel speedup is not linear. so it might actually not be 2x slower but maybe 1.5x or something depending on IPC overhead and all that

I think the quote from IBM means it'd take 2.5 days to run. That doesn't include actually implementing it.

That might be a pretty expensive run on a supercomputer