Lc0 cudnn does not parallelize properly across multiple GPUs

[chara1ampos]

I am trying to run Alexander's lc0 cudnn on 3 GPUs (1 Titan V + 2 x 1080).
I use the command
lc0.exe -w weights.txt --no-smart-pruning --backend=multiplexing "--backend-opts=x(backend=cudnn,gpu=0,max_batch=512),y(backend=cudnn,gpu=1,max_batch=256),z(backend=cudnn,gpu=2,max_batch=256)" --threads=4
and then type
go nodes 130000
to do a benchmark. However, using more GPUs than just the Titan V does not help the NPS get any higher. In fact, NPS gets slightly lower. Also, the utilization of each GPU is 30%, whereas if I run lc0 on one GPU alone, its utilization is 90%.

Why doesn't lc0 properly parallelize / fully utilize the multiple GPUs?