Recently, I placed an order for NVIDIA Jetson Nano development kit, but it will take some time to get it because there is no stock in China. NVIDIA Jetson Nano, a Raspberry Pi on steroids, is a perfect example of NVIDIA Jetson Nano. Raccoons. Be/NVIDIa-Jets… , slightly abridged.

After testing the Google Coral USB accelerator last week, we also reviewed the NVidia Jetson Nano this week.

For the first time out

What’s in the box?

There’s a Jetson Nano in the box… That answer is a little short, so let’s talk about what the Jetson Nano is:

It’s basically NVidia’s answer to Google’s Coral Edge TPU, which aims to run artificial intelligence (EdgeAI) on the Edge. Interestingly, they implement computing power for large models in completely different ways.

Jetson Nano basic specifications: quad-core ARM Cortex-A57, 4GB LPDDR4 RAM, 128CUDA Core Maxwell GPU.

NVIDIA

What’s the difference?

If The Google Coral TPU is an ASIC (dedicated integrated circuit), the NVidia Jetson is an ARM CPU with a 128-core GPU, a more traditional and abstract device.

ASIC, abstract, what?

An abstract device is a device that can do many different things, and it has many uses. In contrast, ASICS were developed to do only one thing. But: Of course, nothing is free, and Jetson, too, carries efficiency costs. Internet data showed Coral scores were much better when MobileNet V2 was used, and the Jetson Nano used more power than Google Coral. But MobileNet V2 is Coral’s only operational network so far. (Because it is an ASIC, it only runs the model developed for it, and custom networks compiled through the Google compiler are also available)

I might do another blog about the comparison between Jetson Nano and Coral Edge TPU, which I think is an interesting story.

Comparison – Source: NVIDIA

Yes, when using MobileNet to detect an object, it was outperformed by Google Coral. But that’s not the point. The point is that it’s a powerful 64-bit ARM with a 128-core GPU. It can run anything. I have installed and tested tensorFlow-Gpus, just like any other DESKTOP system with CUDA-capable NVIDIA Gpus.

An anecdote shows just how powerful it is: Today I forgot about my macbook and decided to work on the Jetson Nano. Yes, the UI part of the operating system seemed a little less responsive than macOS on my old mainframes, but all ML tasks were faster thanks to CUDA acceleration.

The road to test

The Operating system for the Jetson Nano (JetPack, Ubuntu 18.04 LTS) does not appear to be pre-installed with TensorFlow. Of course, most of the required packages can be easily installed via PIP or apt-get.

Correctly identify pikachu classification with GPU

conclusion

The NVIDIA Jetson Nano looks like a good engineering product, small, cool and powerful. I’m pretty sure it will soon be easy to port some AI projects from the desktop, but it’s not as easy to use as Google Coral. But I think it should be compared to the Raspberry PI (as the title suggests) : it’s an RPI-like product! I’m really looking forward to doing some work on it and feel like it’s going to be really sexy when combined with the Coral USB accelerator! I think it’s a shame that TensorFlow isn’t preconfigured, which would make it easier to get started.

I’ll stop here and do some performance testing soon, where I’ll compare the Jetson Nano, Raspberry Pi, RPi + Coral, Jetson Nano + Coral and desktop I7 + GTX1080, using MobileNet V2 to sort the images. Stay tuned!

Source: NVIDIA