Skip to main content
. 2023 Jun 27;13:10391. doi: 10.1038/s41598-023-34436-w

Figure 9.

Figure 9

Inference latency results for (a) SimpleCNN and (b) UNet running on various edge devices connected to an NVIDIA Jetson Nano. We compare models accelerated in the Myriad-X chip, in the Coral EdgeTPU and using CUDA on the NVIDIA Jetson Nano (NVIDIA Maxwell with 128 CUDA cores). Results in Jetson Nano ARM CPU are ommited for UNet as they are an order of magnitude slower. Results are reported as tiles processed per second. While larger tiles may take longer to process, fewer tiles are required for a given image size. Where results for a particular framework and/or device are missing, a model was not available at that tile size due to memory limitations.