I am planning to run two inference logic having two underlying different models. I am evaluating the option of the Google Coral USB accelerator attached to Nano( using USB). With this, I would be running one inference logic on Coral USB and another using CPU. I would be using GPU only to record the video as well as to generate resultant frames. Do you see any issue with going with this approach? I also required to upload inference results in AWS cloud. which would happen through Jetson CPU.
Sorry that we don’t have too much experience as your use case.
But here are some suggestions for you.
First, it’s recommended to check if your second model can reach the expected performance with the Nano CPU (A57).
And it’s recommended to check the bandwidth of Nano and Coral to see if they meet your requirement.
Thanks @AastaLLL for quick response. Can you confirm if there’s any compatibility issues in case if we attach Coral USB Accelerator to Jetson nano. I am more interested in if there’s any impact on performance of Nano because of Coral USB, its connected using USB 3.1 (gen 1) port with max transfer speed of 5Gb/s.
Also, interested to know, would there be any long term impact on Nano device or its power utilisation as Coral USB requires 5V . As per their documentation “The USB Accelerator is powered by 5 V from the USB interface through on-board power management chip (PMIC).The device should be connected to a host computer with a USB cable that provides at least 500 mA at 5 V. Please note that when the device runs at maximum speed, it can draw a peak current of 900 mA.” With peak performance settings, it can able to handle 4 TOPS total peak performance (int8).
My gut feeling is that you’ll have some trouble if you need to actually achieve a consistent USB3 gen. 1 throughput. Also, you’d have to power the Jetson through the barrel jack and not with USB since the micro-USB is barely enough to power a Nano. You could use a powered HUB for the Coral if needed, and perhaps even using the barrel jack will at times require a powered HUB. You’d also have to find out if the accelerator has ARMv-8a/arm64/aarch64 drivers for the 4.9.140 kernel (if it works on RPi 64-bit, then the architecture is good to go, but perhaps not the older kernel).
Just a thought…the Nano dev kit has an m.2 key E. I don’t know if the “A+E” means it can use either, probably it does, but you’d need to verify.
The other thing is that not all m.2 devices will necessarily fit. The single-sided m.2 will, but I don’t know if the Coral device is single-sided or not. If it is mounting components of significant size on both top and bottom, then it might fit, but it could be a problem.
You might also need to find out how much heat this generates. The Nano dev kit m.2 key E slot is sandwiched between the carrier board and the Nano module. Not sure about heat output from the Coral, nor how much that heat might bother either itself or nearby circuitry.