Lab and Facilities
Location: EE168
Sample video:
This sample video shows our object detection stack running on the Jetson TX2 in our testbed. The stack can respond to variation of content characteristics and contention on the device to keep the latency bounded.
Security emphasis:
Instead of using public IPs for every device, we plan to convert the testbed to have a single public authentication gateway for all devices in order to minimize attack vector surface area.
An authentication mechanism, such as Shibboleth will be in place to handle the accounts.
Device performance hierarchy, least to most powerful:
Jetson Nano
Jetson TX2
Jetson Xavier NX
Jetson AGX Xavier
Multitude of devices:
Our IoT testbed contains a variety of popular OEM IoT AI SoCs, available for access to local, and outside researchers
Current setup contains 25 devices
Jetson Nano - 128 Maxwell CUDA Cores
Devices owned: 5
Good for light workloads, best for executing edge-trained AI models. Lowest power consumption among our testbed devices.
Jetson TX2 - 256-core Pascal CUDA Cores
Devices owned: 10
A step up from the Nano.
Jetson Xavier NX - 384 Volta CUDA Cores + 48 Tensor cores
Devices owned: 10
Second fastest Jetson. On-device training performance is orders of magnitude higher than the TX2, higher power consumption.
Jetson AGX Xavier - 512-Core Volta GPU with Tensor Cores
Devices owned: 3
Delivers 20X the performance and 10X the energy efficiency of its predecessor, the NVIDIA Jetson TX2.
How to connect?
ssh username@hostname.ecn.purdue.edu
Account creation, instructions, and troubleshooting support requests can be sent to:
tratkus@purdue.edu
Ideal setup:
