I would say that both observations are correct lims are too big, and 3LP may not be suitable. It's also likely that I=14 is better than A=28; it may be that these even Avalues are only faster in unusual cases (say, a lowyielding poly that would normally be a size to use I=14).
That said, 20% slower for 1 digit larger isn't a huge miss; clearly not faster, but not slowerenough to rule out 3LP as still possibly a good idea. The largerthanyourc167 matrix tells us you didn't oversieve, and we know 3LP matrices turn out bigger.
If we go 2LP on both sides, I suggest also using lambda0 & lambda1 of 1.85 or so to reduce the number of relations needed.
So, possible changes:
I=14
lim's of, say, 50M and 70M
2LP
My notes include a C167 sent to nfs@home 14e queue, in which I used lim's of 67M on both sides. ggnfs likes powerof2 lim's, but that indicates 50/70 are more likely correct than 80/110.
