Native lanes should always give better performance as they're a direct lane to the cpu. PLX lanes add latency and slight performance hits.

LightEVGA
I agree. But if PE1 is going to reduce from x16 to x8 when the second video card goes into PE4 -- what's the difference between having a PE2 that is also native x8 from CPU? Atleast on PE2 it adds another x8 from PLX on top of the x8 from the CPU. You get what I'm saying? 
 
Perhaps you're saying if the PLX adds another x8 on top of the native x8 it'll add latency? 
 
As for the latency, I thought modern PLX such as one in this board have latency that are immeasurable? 
 
Thanks.

Whichever slots run both cards at x8 each is where they need to be. Running through the PLX will add latency which will slow performance a bit. It may be low latency, but it's still added latency that you won't get while running over native lanes.

PLX, much like the older NF200's we used for Gen2 (X58/P55) are a means to an end.  Z170 i5's/i7's have 16 lanes, so some means of replication is needed to run more than 2 way SLI.  Typically on something specifically GPU dependent you see that the X series boards outperform in 3/4 way SLI, this is due to having all 40 lanes on the CPU natively, and not having to use a PLX chip.  PLX will give a non-extreme board the ability to run more video cards than normal, however it is at a cost and the performance will never be as good as it could be through CPU only lanes, but far better than the alternative of just being capped at 2 cards. 
The byproduct of lane replication is latency there is no avoiding this, which is why you have higher bandwidth, but the more lanes you can run through the CPU without a PLX the faster you will run. This was something I worked with the engineering team on to make sure the information was correct when it was released.   They ran a series of benchmarks and while slots 2 and 4 for 2 way SLI provided more throughput ability, the combination of 1 and 4, giving you less total lanes, but more derived form the CPU and consistently tested faster than 16/16 through PLX.