20 nm is an option but it's a half node so it's only about 2x as good as 28 nm.
2x? think you need to look up your scaling numbers for node shrinks (especially theoretical scaling vs actual on the latest generations), also 28nm is a half node so the jump is similar as for example 32 > 22nm.
"Half node step" :-)
20nm should give around 2x- it will depend on the exact process of course but there's an increase in density and an uplift in frequency (or reduction in power consumption). We're talking about a class if design where trivial parallelism leads to almost linear speedups.