I'll start by saying I have always been an advocate of a dedicated PhysX card in the past even though at times they made very little gain's but still gains were made and games I play a lot such as Borderlands series and Metro use physx.
 
I have had the following graphic card configs:
 
GTX 580 main Gtx 460 PhysX= Great results.
Gtx 680 main Gtx 460 PhysX= Great results.
Gtx 780 Main Gtx 460 PhysX= Great results.
Gtx 980TI Main Gtx 780 Physx= Great Results.
 
And now...
Gtx 1080 Main and any PhysX benchmark as well as games that use PhysX my fps and benchmark score went lower when using a Gtx 780 for a dedicated PhysX card.
 
I do realize the GTX 1080 is a more powerfull card But previously I had a GTX 980ti Kingpin overclocked to almost 1500mhz and the Gtx 780 used for PhysX in this config helped me a lot and gtx 1080 comes in and does everything all by itself better then with a PhysX card installed.
 
I figure this would be an interesting discussion because I still see people ask this question about dedicated PhysX cards in conjunction with Gtx 1070, Gtx 1080, Gtx 1080ti tier card's and I still see the same responses such as not worth it,creates to much heat,not enough games use it etc.. but to me even small gaines make it worth it if you have a card to use for PhysX just laying around collecting dust..but something has changed starting with the 10 series cards and my opinion about using a dedicated PhysX card has also changed.
 
I remember Nvidia showing some type of technology they were hypeing up with the release of the Gtx 1080 which at the time I thought was bit of well just hype to sell the cards but I now think there is something to this. Maybe someone much more educated then myself can explain this one to the uneducated (me) :)
 
The card sure is an extremely powerfull quiet and efficient design.

That's interesting, my guess is that it's possibly a latency issue.

I don't know the flow path from CPU to the physx card (does it have to go to GPU first or not), but at the very least, once the physx gets computed, it has to travel to the GPU for objects to be drawn. That is already an extra "step."

Overall, it's not really an actual step if you consider that it does the same on board the 1080 (aka primary card) if you think about the CUDA cores on the card as a "separate" unit. The difference then becomes travel time (aka latency) from CUDA core on 1080 vs 980ti to main gpraphics core on the 1080.

If the main core is fast enough, that time difference eventually becomes a "bottleneck" of sorts. So, maybe the 1080 has just enough processing power for you set of tests to cross that threshold, going from alleviating the GPU to holdong it back.

Did you monitor usages on both primary and physx only cards?

I had a little bit of fun with Batman Arkham Knight on Titan X Maxwells and a 970. https://forums.geforce.co...bit-of-batman-ak-fun-/

Edit: oops, I thought you were using the 980ti and not the 780 as the physx card. So yeah, I think it's the 780 itself holding you back.

Here is my 2 cents on PhysX.  If your game/bench is PhysX ready, this is mainly ad games though, and the card/cards you run as main GPUs support PhysX, look at your usage, if your GPU is 70-75% usage and below on average, and you arent seeing major spikes in usage during explosions of pother PhysX heavy load, then a PhysX card is not going to really help, as the main GPU is already doing it, and keeping its utilization in check.  Next, if you do see room for one, then the next thing is what is the secondary card, in related performance to the main card(s), if it is more than 1 generation back (900 series, not necessarily meaning TI cards as their own series) it is likely you will not see very much gains, and if you force a PhysX load to a card that cannot process PhysX as fast as the main card with its other load, which a 700 series card may not, then it may well work against you and slow you down some.

Check out the 3 minute mark and use it as a guide to see if your 780 is holding back your 1080: https://www.youtube.com/watch?v=H9nZWEekm9c