RPI vs Jetson nano

Make it sense to use a Jetson Nano for DonkeyCar ?

Are there any performance measurements RPI vs Jetson Nano available ?

Short answer: Yes !!

Longer answer:

One of the reasons why the Jetson Nano is very exciting for us is that it has a lot more headroom for inference. Right now we can do 20fps for inference if we use a Raspberry Pi 3B+ class device.

However, having experimented with deeper neural nets - this will be a bottleneck (inference happens on the CPU for the Pi). The Jetson Nano has 128 CUDA cores, which allows us to do so much more.

Keep in mind:

  • It’s early days with the Jetson Nano.
    So the level of support for the Nano is not as mature as the RPi 3.
  • You should be a little bit familiar with the donkey source. You will probably have to make tweaks for your setup until we standardize things for the Jetson Nano.
2 Likes

New RPI4 is out.

I’m very interested in new performance benchmarks.

RPI4 vs Nano

1 Like

One issue with Nano I’ve run into is that the generic raspberry pi camera are not compatible with Nano. You must use the newer v2 camera. I’ve not found a cheap source for a v2 camera with wide angle lens. (There are some options available, but they’re close to $60 USD for camera + wide lens)

If somebody has found a Nano Compatible pi camera with wide lens that can be purchased < $25 USD, please let me know.

Replying rather late on this - but there are now fairly cheap wide-angle camera modules for the Nano. Waveshare have a whole range - this is the 160° module at $23: https://www.waveshare.com/imx219-160-camera.htm But they also have 200°, IR etc.

If you’re not familiar with Waveshare, they’re the company that basically provide nearly all the third party cameras modules you see for the Raspberry Pi, i.e. the ones that aren’t officially blessed by the Pi Foundation (and produced by RS etc.). If you look closely at most of the modules sold on Amazon and eBay you’ll see they’re Waveshare even if the brand name used in the product description is something else (for whatever reason Chinese manufacturers often seem happy to let downstream resellers rebrand their products even when they’re relatively big like Waveshare).

When the Nano came out, I just replaced the module on a genuine v2 Raspberry Pi camera with a $12 wide angle module from Waveshare: https://www.waveshare.com/product/modules/cameras/imx219-d160.htm This is a cheaper option if you already have a v2 Raspberry Pi camera.

1 Like

Here is one on Amazon for $25. I had to change one line in the camera.py file to flip the image.

Yes, the v2 cameras can be had for ~$25. However they are not wide FOV which is highly desirable on this project. You then end up having to spend more to get the wide FOV lens. ~$20 (Given that the Nano is $99, it starts to get ridiculous spending ~$45 just for the camera itself. The old v1 cameras with wide FOVs were < ~$20 total.)

It looks like what @ghawkins found is probably the best deal going in terms of wide FOVs. I’m not sure that was available when I first started looking.

It also looks like Waveshare has a nice IR wide FOV lens as well. That might be interesting. As would a cheap-isa stereo camera (Which I didn’t see listed.)

store.donkeycar.com has them now for $27 shipped. https://store.donkeycar.com/products/nvidia-jetson-camera-for-donkey

1 Like