I have three 4mp cameras running 25fps and at 720p video decoding is set to 'passthrough' and hardware encoding is 'cuda'. Cuda is installed with the non free nvidia driver for use with zmeventnotification server and ML hooks.
The most obvious explanation is in the Buffers Tab, for 'Image Buffer Size' it is suggested to run 3-5 frames only, I had been running much more(wiki still says no more than 50) in my previous build. Why the change? My post and pre event buffers are set to 200 so there should be something in the buffer que during an event.
Less obvious is that I went from 1GB of VRAM to 8GB, are the camera frames being buffered in the gpu instead of /dev/shm now?
Maybe its just a reporting error? like I said it seems to run fine, much better with a real gpu thats for sure.
