Importantly, I have 12 GPUs in every node, with an i3-7100 cpu. The nodes with gtx1070ti keep the CPU at 100% utilisation (on both 1.45 and 1.46/1.47) and the GPUs cannot handle anything else (even nvidia-smi fails, takes 5 minutes to show 1 line).
Try
--watchdog 0 --pec 0 options to temper gminer's thirst for NVML calls. And, I guess, you have old nvidia drivers like 3xx.something. Try 4xx branch, it's kinda optimized for multi-gpu environment--less LA, more responsiveness.
p.s. @Zminer777, it would be nice to add smth like
--no-nvml option, it'll help a lot.