one thing to note is that the gpu hashcat uses a very complex cacheing, which basically boils down to a buffer per password length, i.e. a buffer for 12 character passwords, a buffer for 13 character passwords, and shoves passwords in accordingly. Once a password buffer hits a certain threshold, they all get sent to be cracked. This means you may not get feedback for a certain length try for a while, if it's not a common length. One thing to consider would be to figure out that threshold and keep things internally until you hit it, so you can fill a buffer in one shot. Realistically, you aren't going to get around this very easily, as this buffering is done for performance reasons, in that some of the kernels used rely on same length passwords for performance increases.
Another thing to keep in mind is that hashcat can generally test guesses faster then you can feed them over the various connections, i.e. memory -> memory copy for pipe memory -> pcie -> card for loading to the GPU. Also, in cracking rigs, you often have multiple cards over pcie x1 connections, further lowering bandwidth for those cards. One thing I'd highly recommend is trying it with a rules. This allows the GPU to expand your word guesses out on the GPU itself, turning a single guess into thousands of guesses or more. This bypasses memory bandwidth issues, as it's done on the GPU itself, and also means that a lot of rather trivial checks can happen automatically (try it with a 1 at the end, try it with a 2 at the end, and so forth).
