Jetson Xavier NX versus Jetson Nano - Benchmarks and Custom Model Inference

Link to article:
www.hackster.io/dmitrywat/jet...
The hardware for this video was kindly provided by Seeed studio, check out Nvidia Jetson Xavier NX, Nvidia Jetson Nano and other hardware at Seeed studio store!
www.seeedstudio.com/NVIDIA-Je...
www.seeedstudio.com/NVIDIA-Je...
Also check out reComputer Jetson series - compact edge computers built with NVIDIA advanced AI embedded systems: Jetson-10 (Nano) and Jetson-20 (Xavier NX):
wiki.seeedstudio.com/reComput...
0:00 Introduction
2:32 Hardware
5:43 Specs
8:41 Benchmarking
10:32 Container demo
13:15 Custom model inference
15:27 Conclusion
Add me on LinkedIn if you have any questions:
/ dmitry-maslov-ai
Credits for the music:
Daft Punk - Get lucky (Incredible Apollo chillout mix)
• Daft Punk - Get lucky ...

[@SkyHorseTech]

Verdict: If thinking about the Tx2 the Xavier NX is the only answer. Thank you for great explanation and details.

[@Hardwareai]

Oh, absolutely, with NX on the market, there's no reason to go for TX2. Thank you for your kind words!

[@glennmura]

Great video, now I know what to expect "if" I upgrade from my Nano!

[@Hardwareai]

Thanks! What are you using Nano for now?

[@JBoy340a]

Nice work. Just got my xavier nx and this video gave me a lot of ideas. I have a NVMe SSD on it so maybe I can retry the cars benchmark.

[@Hardwareai]

Yes, great idea! Would be interesting to see if it gives performance improvement

[@adelhammouda6563]

Thanks Its so useful for my project

[@Hardwareai]

Great 👍

[@anpr]

Excellent

[@Hardwareai]

Thank you so much 😀

[@carlosgaonad]

Hi, I want to run 2 ObjectDetection models simultaneously on a video stream (A traditional YOLO and a customized mask-detector) & I need to do this for around 5-8 IP security cameras.Will the JetsonXavier do the job? How easy is it to deploy and optimize a customized model on it?

[@Hardwareai]

Not easy, I would say, but certainly possible. For using multiple video streams you will certainly need to familiarize yourself with DeepStream to avoid bottlenecks when getting the data from cameras. Traditional YOLO runs flawlessly on Jetsons, for your customized mask detector, how well will it run on Jetson depends on the architecture.

[@hidlayer-videoanalyticscom3479]

Jetson Nano can work with 6-12 streams but need to spend a lot of time to achieve that kind of performance.

[@levbereggelezo]

Very good

[@Hardwareai]

Thank you for your continued support!

[@TechGuru-kl1yw]

Thank you..

[@Hardwareai]

Funny thing, I have made this video twice xD The first time I recorded all the videos, almost finished editing when I found out that I mispronounced the word "Xavier" every single time - I pronounced it as "ks-avier" instead of "ex-avier"... So I had to record and edit again

[@curtishorn1267]

Have you tried putting the nano in the NX so it can have access to the M.2 drive?

[@Hardwareai]

No, I haven't tried that yet. Actually my Nano module is old revision(A-02), so I don't think it will work with NX carrier board.

[@roshanraj999]

Just wondering, what is the physical weight of Nano vs Xavier NX with the carrier board.

[@Hardwareai]

I just measured with electronic scale for you - Jetson Nano 143 grams, Xavier NX 177 grams.

[@eepdoesstuff]

What would you suggest for someone on a tight budget that is trying to train AI to perform actions based on words it hears? I have plans to work on something that will listen for key words during live streams and then perform certain actions. I'll probably only have up to about $1700 USD (or $2200 aud) and I'm wanting to make this my entry into machine learning/programming.

[@Hardwareai]

Jetson Nano sounds suitable. Since you're only going to be looking for keywords and not transcribing the speech, you might get away with Raspberry Pi or even something simpler. Check out my latest video on speech recognition on microcontrollers by the way kzbin.info/www/bejne/eYfUZZabqZqdnas

[@252.studio7]

older video but what if you're not wanting to run an HD camera but a depth camera x 2 or 3 would the NANO be ok to do that? at this time the Xavier cost $1,000 and up or around that

[@Hardwareai]

Oh-oh. The chip shortages? You mean 2 or 3 depth cameras simultaneously? That would not work on Nano. I mean just getting video feed should work, but running inference would be too slow. For depth images, have you considered OAK kits? Like an OAK-D?

[@dajmispokojchuju8556]

9:33 why is this pin bent? Should there be a joint for these two pins? Mean this one on the forefront, which is probably touching the one behind

[@Hardwareai]

Good eyes! Yes, the pin is bent, but they're not touching. I took out the board from the drawer to double check just now :)

[@rockroborocket7342]

Great one... I liked the way... I had one doubt. Can I ask able to use NCS2 ( Nvidia Compute stick 2 ) with Jetson nano? Is it possible? Is it will increase the fps on image processing?

[@Hardwareai]

Of course. You can offload some of the neural network inference to NCS2 - in case you're running multiple networks. But the power consumption will be terrible.

[@akhileshagnihotri1378]

Please tell me Jetson Xavier nx can be good option for full time use as personal computer for me . I am a software engineer and iot devloper

[@TECHNDJ]

Not much tho.. I still had issues with even basic coding other than AI applications. Will recommend a discrete PC and this one for more reliable productivity.

[@Hardwareai]

I wouldn't recommend using Xavier NX as replacement for personal computer, at least it's not something that I would do myself. It's running a fork of Ubuntu by NVIDIA and has ARM architecture, so if do encounter any quirks or difficulties that you are not able to resolve on your own, your only hope is Nvidia forums :) Best leave it to what it does the best - AI and CV, as an embedded device in robots and portable electronics.

[@mbuhamad]

Hello, I just got Jetson Nano. I'm thinking to make car number plate recognition. Would you please point me to source of information which can help me to accomplish this.

[@Hardwareai]

License plate recognition is not an easy task - you can have a look at some other articles, for example this one medium.com/@theophilebuyssens/license-plate-recognition-using-opencv-yolo-and-keras-f5bfe03afc65 This one uses multi-stage approach - a)detect license plate, crop the image 2)detect characters, crop the image 3)recognize the characters in cropped images All of this should be possible with Jetson Nano. In near future I plan to research more about using TensorRT and make another video on model optimization with it!

[@mbuhamad]

@@Hardwareai Thank you so much for your response. I have good knowledge for SBC, Linux, and python. However, I'm totally new to AI. I have a lot to learn. Can't wait to more of your videos. Thank you again

[@nazaninfasihi1400]

hello my jetson nano doesn't work .for configuration , when i connect it to the power supply the green light is ok .but the USB ports do not work and nothing appears on the monitor .can you help me to figure it out?

[@utubenamingsucks]

how do you test the USB port and how do you know it doesn't work? Have you tried to flash the sdcard rightaway? download the jetpack image from nvidia website and make sure it is the correct image (i think jetson nano 2GB and 4GB nano has different jetpack)

[@Hardwareai]

It is hardly an appropriate place to ask general questions about Jetson Nano :) Perhaps you can try forums.developer.nvidia.com ?

[@alexandrunistoroiu452]

Do you maybe know if blender rendering would take full advantage of xavier? But of nano?

[@Hardwareai]

I don't think it is intended way to use Xavier xD but then again people run Doom on washing machines

[@nanogear164]

Hello, how to identify if a Jetson Nano is B01 version or another? I need a B01 but I don’t know how to identify them

[@Hardwareai]

There is a picture with the differences in the video kzbin.info/www/bejne/qWHPiGZ-oMirY80 basically the easiest way is to count CSI camera connectors: B01has two and older A02 only has one. The one you see in the video is A02 - which means it is NOT compatible with Xavier NX module

[@nanogear164]

Hardware.ai ohh thank you...

[@naturepi]

Can you upgrade Ubuntu to 20.04 and run GPU accelerated content rendering in Blender or Kdenlive? Have you try a basic video editing on these units? Thanks

[@Hardwareai]

Hm... Not really. I am not really into 3D design or animation :)

[@naturepi]

@@Hardwareai too bad...anybody else?

[@krukhlis]

Well, I know for sure that all FFMPEG-based video encoding products will not be able to use GPU acceleration of Jetson Nano ( and as far as I know X NX as well). NVidia provides hardware media accelated video encoding/decoding only through the L4T Accelerated GStreamer API. I can't find anything recent about Blender support for L4T GStreamer on Jetson boards. Regardless -- software only (CPU based) FFMPEG works fine.

[@krukhlis]

@@naturepi Also, watch this( even considering it's related to Jetson Nano only) kzbin.info/www/bejne/ppm3dImijsaod8U&feature=emb_logo

[@naturepi]

@@krukhlis thanks for update, i still would like to see someone of "reviewers" to post the test. Latest Kdenlive now supports parrallel CPU based processing, which cuts the time of encoding by X times(3-4 in my case) and since nVidia positioning this unit as AI video streamer, they should have built in native hardware instructions for encodering/decodering process, including h265 support. I know arm chips only support h265 at some level. We need to see more tests, not just repeted marketing demos. Anybody?

[@sahilpawaskar5512]

Is it possible some how that macbook Air M1 can be use an monitor for Jetson board ? Please reply I am new to this field.

[@Hardwareai]

You can set up a VNC server, if that's what you're asking. You'll still need to do initial set up though - if you don't have a screen, you can use USB-to-serial converter to do it over Serial port.

[@sahilpawaskar5512]

@@Hardwareai Thank you so much for replying and providing the solution, I will definitely work on it. 👍👍👍

[@smitsavaliya]

Hi, till now i am using raspberry pi 4 ( 4gb ram and afterward 8gb ram ) but it is not performing well on my project Autonomous Vehicle, and i am confused at buying jetson nano and xavier nx, which should i buy for Autonomous Vehicle Project.

[@smitsavaliya]

my work is image processing using computer vision, object detection , image segmentation, controlling hardware , streaming all the real time data to mobile/tablet via an app which i have make and artificial intelligence, machine learning . A short of making Tesla Autopilot mini version.

[@Hardwareai]

Yeah, poor Raspberry Pi might go in flames at your workloads xD Surely Xavier NX. For prototype at least

[@smitsavaliya]

@@Hardwareai Thank You for reply

[@t-lm]

is there already a project on it for multicam NVR with object detection?

[@Hardwareai]

Hmm, I would think so. There are examples available for sure and pre-trained models for cars, people, etc. If you need to train your own model you can use nvidia transfer learning toolkit or aXeleRate

[@hidlayer-videoanalyticscom3479]

Yep, we already have. If have interest can see it on our channel.

[@dachosenone366]

Question could these chips be used in game console if nvidia was to mass production the chip!

[@Hardwareai]

They probably won't be xD too many things that gaming console doesn't need, for example above mentioned DLA engines

[@dachosenone366]

@@Hardwareai Do you know what chips or board I get for android testing and gaming.

[@krukhlis]

@@dachosenone366 You can buy Nvidia Shield and play games( even retro/emulation) there. It's super powerful even considering it's a grandpa of these boards. Regarding Jetson Xavier NX -- it's really capable of running emulators even for different PlayStations in >30FPS mode. But as @Hardware.ai has replied -- why should you pay 400 USD and use it only for 1/4 of its capabilities? Jetson Nano is also very capable in emulation and costs

[@ehrazimam7967]

Can i use Xavier NX developer kit with nano module

[@Hardwareai]

From what I remember - yes. As long it's a newer rev. of Jetson Nano module.

[@soorajsurendran648]

Can we interface 6 camera to jetson Xavier .

[@Hardwareai]

Yes. From datasheet here developer.nvidia.com/embedded/jetson-xavier-nx

[@shivammathur1583]

I love jetson but I want to use it for robotics and AI applications along with ML I have a raspberry pi 4 which is the best option ?

[@marcvanzyl8556]

For robotics - similar. For AI - that is basically what the jetson was meant for, so def Jetson

[@shivammathur1583]

@@marcvanzyl8556 But which one or is google coral board

[@imagidan]

@@shivammathur1583 What is the thoughest ai task you want your board to do? How fast do you want it to do the task?

[@shivammathur1583]

I want it to use vision learning in 20 seconds and do robotics and hi speed computing something like quantum computing at a small scale

[@imagidan]

@@shivammathur1583 I think quantum computing aside, both boards could do this. Obviously, Xavier NX is faster at vision tasks and in general, but you would have to look at the projects being developed with both and see what is best for you. For me, for example, I want to make an autonomous drone and I used the skydio r2 (running jetson tx2) as a comparison. That is why I chose the xavier nx.

[@denvera1g1]

I want the dev boaord for NX for my nano

[@Hardwareai]

I think the carrier boards are compatible between NX developer kit and Jetson Nano B01 module/devkit.

[@kobki66]

Hi. Do you think It's possible to mine crypto on Jetson?

[@Hardwareai]

Well, theoretically it is possible to mine crypto on a calculator... the question is - is it feasible (time and money-wise). The answer for both - no.

[@Jkauppa]

simple install for android, thanks

[@Hardwareai]

pardon me?

[@Jkauppa]

@@Hardwareai port lineageOS to nvidia sbc, similar to konstakang lineageOS port to raspberry pi

[@Tau-qr7f]

the audio seems more natural at speed 1.5x 😅

[@Hardwareai]

Erm xD Anyways, feel free to watch at any speed you like

[@SaimasCookingSecrets]

@@Hardwareai Can you please guide if I can use TX2 NX SOM on nano developer kit?

[@tenminutetokyo2643]

DOOD!

[@Hardwareai]

Does that mean good? Or? xD

[@tenminutetokyo2643]

@@Hardwareai It’s a Silicon Valley thing.

[@whothefoxcares]

Nvidia could sell a baller developer system with a nice case, but it wants you to suffer. *IMHO*

[@Hardwareai]

Of course they are :) have you read their documentation?

[@hidlayer-videoanalyticscom3479]

We build software that works good with Jetson boards and make use of this board easy to use. Cases can see on our channel.

[@flytie3861]

1.25x your welcome

[@Hardwareai]

You mean the playback speed? ;) Do I speak that slow?

[@thunderbolt997]

@fly tie if he would be speaking faster you would be complaining about unclear words

[@MrLostGravity]

@@HardwareaiAt times, yes