DR render times longer?

[biochemical_animations]

I'm setting up my small render farm with GPUs and running tests to make sure everything is working as expected.

My WS has 2x980Ti's and when I render a scene I have with a refractive object i get a render time of 3:29s noise at 0.05.
When I add a render node with a single 970 to the mix via DR, my render time is 4:03s...an increase in render time.

I noticed the light cache gets built twice for the DR render, which took 30s approx, but why would the 3 GPUs not still render faster?

I'm using 3.5 beta on all, max 2016, win 7 64 bit pro on all. Only asset is an hdri on a dome light that I have copied to the node under the same path as the ws.

Any idea what I could check here?

This is using production mode RT GPU...thanks.

[biochemical_animations]

Maybe I'm getting a bottleneck somewhere, perhaps the network?

If I ran four gpus from the same mb, would this help?

Are their limitations to running multiple gpus past a certain point...I just read another post here where it was mentioned 2 cards might be faster then 3, if ui performance causes reducing rpp and rpb. I'm still trying to figure out what those are and how they work.