Here are some numbers illustrating the performance I have seen for the Nvidia Jetson Nano so far.
Training this PyTorch Image Classifier with the flower data set, took just under 10 hours for the full 30 epochs. Not at all bad, and really quite nice if you consider the average power consumption of less than 8 watts.
Numbers from a workstation have been included for comparison, of course, that one spent more than 500 watts on average both for CPU and GPU training.
Nvidia Jetson Nano:
– GPU (Nvidia Maxwell): < 10 hours
Dell Precision Workstation T7500:
– CPU (Dual Intel Xeon E5620): < 30 hours
– GPU (Nvidia GeForce GTX1060): < 40 minutes
The Jetson Nano clocked in at 857 Mbps as measured with iperf to a Dell Precision T7500 workstation, interconnected with a Netgear ProSAFE GS108T switch. I believe the workstation would do close to 950 Mbps when connected to a proper Cisco Catalyst like the 2960-X.
As one might expect, compared to a Raspberry Pi the Jetson Nano did absolutely beautifully.
Nvidia Jetson Nano: 857 Mbps
Raspberry Pi 3 B+: 292 Mbps
Broadcom NetXtreme BCM5761: 878 Mbps (between two T7500 workstations)