Originally Posted by
Blameless 
Multi-GPU rendering is inherently problematic for latency sensitive games and people because the only rendering mode commonly in use is AFR which requires a queue of frames at least as deep as the number of GPUs involved. This is one of the reasons I've categorically refused to build quad GPU setups...there is simply no way the performance gain from the last few cards involved can overcome the latency penalty of needing to render four or five frames before you are allowed to see the the first. The irony is that older methods, like split frame rendering, supertiling, or even scan-line interleave, didn't have this latency issue, or any microstuttering problems, but were rapidly depreciated in favor of AFR for compatibility and performance scaling reasons.
Anyway, if you haven't done so already, try reducing flipqueuesize to 2, and do whatever else is within your power to reduce any latency that may be introduced. Turn off vsync if you can tolerate tearing, use the highest polling rate possible on your input devices, make sure there is no GPU scaling going on, find out which input your display applies the least processing to and use that, etc.
You may also want to try disabling frame pacing, as it can introduce some latency in some cases. If the increase in microsuttering is not perceptible in the title in question, but the reduction in latency is...
Since you first noticed the issue in a CPU intensive situation (I'd guess the CPU is starting to have issues keeping the present queue full), finding out what settings in game are most CPU dependent, and reducing them, may also help. Shadows and physics are often heavily CPU dependent.
