I use a NAS to store the output and I discovered that each time the network interface is under load the GPU utilisation goes to zero.
This becomes significant when using the ProRes format due to the large amount of data to be written each frame.
Please refactor the file i/o to run on asynchronously to the main thread. I think this can be archived by adding a buffer that holds the previous frame(s) in memory and runs concurrently with other tasks.