**5. Experimental Results**

In this section, we show the results for all the experiments discussed in Section 4 (varying number of devices, free and locked input layer, and all the approaches) for the inference rate maximization objective function. After that, we show the pipeline parallelism factor for each setup to compare the performance of a single device to the distribution performance. Finally, the results of the inference rate maximization are plotted along with results for communication reduction to see how optimizing for one objective function affects the other. Our approaches (greedy algorithm, iRgreedy approach, DN2PCIoT 30R, and DN2PCIoT after the other approaches) were compared to two literature approaches: the per-layers approach (equivalent to popular ML frameworks such as TensorFlow, DIANNE, and DeepX) and METIS. We implemented DN2PCIoT using C++ and executed the experiments on Linux-based operating systems.
