Thanks for sharing the experiment. I'll pass it along to our engineering team, for them to look into. There's not a lot of controlled variables in your testing so drawing conclusions and saying it's an architecture issue seems rather premature, if not misleading. That chart shows a 16% advantage for Haswell over Skylake. Time for pitchforks about the regression of Skylake vs Haswell? I kid, kinda, it's a good area for AMD to look into, but I'm not sure this testing is a "smoking gun" as it were.
We had users with NVidia cards supply results, but their driver has an optimization that only applies for scenes calling the exact same draw call, with no changes, over and over (i.e, floating rocks with no materials, shadows, lights, parallax maps, etc).
We also found that Windows 10 had a performance increase for draw calls over Windows 7. It's a minor gain, a few percentage points, but it's still there.
That Haswell dual core result is an anomaly, I should have not put it in, admittedly. But if you look at the quad core Haswell result, it's much more in-line with what was expected. Skylake made quite the jump in draw call performance over Haswell.