Ok... my own graphs show that the area 'over' the green curve is less than half the area under the 'blue' curve which would seem to indicate the principle you are referring to.
Namely, that reducing memory by 50% only hurts performance by ~20% (eyeballing it). However, it does hurt performance.
I calculated the area under the curve for a 1/1000 reduction in memory requirement (down to the size of SCRYPT) and performance is 253.58x slower...
To get back to equal performance with single threading you would have to go parallel 254 threads each using 1 MB of memory for a total of 253 MB of memory, a 75% memory reduction.
To run this on a GPU the most parallel threads you could get going on a graphics card your would need 2GB video memory. The most CPU-core equivalent performance you could get is 8 x.
All of this assumes of course that the GPU threads are of equal performance to the CPU threads (which they are not). So an 8-core CPU is still in the same neighborhood has a 2GB GPU using your trick to reduce memory usage.