*4.8. DN*2*PCIoT 30R*

The fifth approach that we used for the experiments was the application of DN2PCIoT starting from random-generated partitionings. This approach executed DN2PCIoT 30 times starting from different random-generated partitionings and we report the best value achieved in these 30 executions. It is worth noting that this approach was only executed for the LeNet 2:1 model due to the more costly execution of LeNet 1:1 starting from a random partitioning. This was the only approach that did not employ LeNet 1:1. Furthermore, DN2PCIoT can discard some devices when they are not necessary, i.e., if DN2PCIoT finds a better partitioning with fewer devices.
