We need your like! Please help me. The more likes we will have, the more views will be. The more views gives more subscribers. The more subscribers we will have, the more videos we will make. And we hope these these videos will be useful for you.
It's a problem of the TH-cam algorithm™. I have my channel for 12 years now and still only 500 subscribers, despite lots of good videos (based on the like/dislike ratio). TH-cam promotes only clickbait crap but never small TH-camrs, especially in the science category.
As a robot idea, you may want to build a rc self-driving car with depth estimation from mono camera deeplearning models, like monodepth or something more recent. I know this is not the type of robots you usually build but it could be cool. Or you could build a robot arm which tracks an object class and tries to grab it if it is close enough (depth and pose estimation)
These might be useful for an idea I've had where you had a bunch of these in an array out in a forest to feed an MR AR experience. Attach lidar sensor and feed headsets with proper positional data. First app is forest digital paintball.
I'm really surprised by the amount of memory these things have. If they can cram 64GB @205GB/s in such a small module, or 16GB at 102GB/s on a SO-DIMM stick, why don't they do this with their desktop GPUs as well? You can never have too much video memory :) Also, if you think about it, the Jetson AGX Orin are not too different from the Apple M1 Mac Mini.
@@AltMarc I know; I use the RTX 8000s at JURECA-DC for my work. Still their ~$7k price tag seems way off. The memory chips cost a lot less. Hardware vendors could put much more memory on their GPUs and still be profitable. Right now, compute power rising much faster than memory capacity/bandwidth creates big issues for software. The benefit of more/faster memory for consumers, video producers and especially scientists would be huge.
Sir, thanks for your share. I've got the same kit and set up the kit with NVidia SDK Manager (from host pc). System works fine, CUDA 11.4 folder is exist in usr/local but CUDA doesn't work even with my basic Opencv-Computer Vision projects. Is that normal ? So, am I suppossed to set up CUDA manually or with host PC
The jetson agx Orin, Can you test stable diffusion. And also a good video editing program like di Vinci. The ai suggest that you can run both on there. I have an old laptop and where as I’m not concerned with playing high end games it looks like this may be a good purchase to consider LLM running locally, stable diffusion, and general video editing 4k video for example.
I am working on a robot platform from scratch and god damn the moment i have a slight excuse to buy something like this i am pulling the trigger but i am not quite there
The speech recognition demo is quite cool, but it would be good if you test it in a noisy (outdoor?) environment to check if the model is robust and usable in real case scenarios really. I like when you conclude by saying "so far it works great" and it recognized "so fireworks..." haha, those models always screw up during those type of moments.
Agreed, though I guess that many companies who want to deploy the less expensive Jetson boards might still buy the Orin units for their developers to use at their desks.
Very interesting thank you. Depth and rgb camera should be included. I am sure it is comming as well as an extensive AI model to recognize common objects and escenarios like animals, trees and stairs. Finally it will be a general purpoise brain. And cheaper I hope jeje. Thank you
I would be curious to see it being used in transfer learning projects, where models based on RGB input are retrained to use thermal IR cameras, e.g. to track and recognise faces in the dark. Or use it in sensor fusion between lasers scanners and (stereo) cameras.
Really depends on the model but since transformers are memory hungry, but this comes in 64GB shared vram memory (as far as I understood it), it should not be a problem as long you be mindful to not over max the available tensor (i.e. 4K video + multiple IR sensors etc). Thats inference. The second part regarding training I don’t understand - training a transformer model at the edge (something that is usually done distributed along many many machines, also because you need tons of training data) is not what I’d say this is good for. Aka Inference with large models: Yes. Training: No.
you use existing models. it would be nice to create and train an own model, for example to detect cupcake and donuts. here you need a lot of computing power ... and if you can do this, you can delop unique applications. maybe it's nonsense what I'm saying here, in any case, many thanks
What you saying is not a nonsense. But it is complicated to make and train the model from zero. First you would need a lot of data, super powerful computer and knowledge of deep learning. That is why TAO toolkit is great. It allows to reuse existing models relatively easy.
Doing all this within 15w to 40w of power is truly amazing. My computer is much faster, but it consumes up to 1600watts of power. Not something you want to throw into a robot unless the robot is on a solar-powered golf-cart with a generator in the back seat, as a backup. Having all these "tuned" and "specialized" AI and CV components is the biggest advantage. Even if it hits 40watts, in the long-term it is consuming only a fraction of the same power as a typical CPU setup would consume, to process the same data. Doing it faster is a real nice bonus to the package. It will not be long before we see a Jetson with the processing power of a 1080 or 2080 GPU, wrapped up in a nice 60watt package, for under $1000 in a kit, I am sure.
Great question. First, for your RTX you need to add motherboard, memory, and CPU, this will be another 700$ at least (12core CPU+32G of memory). So your total cost will be more than 1k$. And your system will be way more bulky (Jetson module is only 100x87mm). Also, your PC will consume around 500W of power. For mobile robot this is a lot (only 1hour on the standard bicycle battery), Jetson consume only 60W at maximum, so it can run 8hours on a standard bicycle battery. Plus Jetson has some additional features: special port for CSI cameras (it is better than USB camera, as it requires less CPU to run); GPIO pins for sensors and other hardware. Also there is at least one practical point: on mobile robots you can have some vibrations (due to the moving), and PCI-E slot does not works properly with vibrations. I know a company, which had a lot of trouble with this.
@@Skyentific "PCI-E slot does not works properly with vibrations" -- a very interesting datapoint, thank you. I wonder if you can share any additional info about this failure mode.
This thing is _$2000_ and is the same size as an ITX PC, and has very limited software support and ecosystem compared to regular compute platforms, so lots of fiddling with compilers and large chance this just kills your project. I kind of don't see the point of this thing, it's compute per $ isn't even competitive with a regular ITX PC with a mid-range GPU from 5 years ago, so what is the point? Maybe it's slightly less power hungry?
2400€? hahahha To expensive for what it give. With normal PC can do more than that developed kit. and is only 4-5 times biger but i sure i dont have problems with the heat disipation.
![เจ็บที่ยังฮัก - ม่อน วรวิทย์ [ Official Mv ] จอนนี่มิวสิค](http://i.ytimg.com/vi/KALIGSRSjFs/mqdefault.jpg)
We need your like! Please help me. The more likes we will have, the more views will be. The more views gives more subscribers. The more subscribers we will have, the more videos we will make. And we hope these these videos will be useful for you.
This thing is truly a monster but it would be nice if you could stick to the more accessible Jetson Nano for most simple projects
It's a problem of the TH-cam algorithm™. I have my channel for 12 years now and still only 500 subscribers, despite lots of good videos (based on the like/dislike ratio). TH-cam promotes only clickbait crap but never small TH-camrs, especially in the science category.
Absolutely amazing! I need to get myself one of those
As a robot idea, you may want to build a rc self-driving car with depth estimation from mono camera deeplearning models, like monodepth or something more recent. I know this is not the type of robots you usually build but it could be cool. Or you could build a robot arm which tracks an object class and tries to grab it if it is close enough (depth and pose estimation)
The true powerhouse here is you! (The jetson AGX Orix is pretty cool too.)
Kind of interesting you can see the attention process of the speech transformer model in action during translation.
Cool! Excelent content. Thank you for sharing.
Thanks for a great video. I would like to see more on how to program it. Specifically, more details on what you did to recognize helmets.
These might be useful for an idea I've had where you had a bunch of these in an array out in a forest to feed an MR AR experience. Attach lidar sensor and feed headsets with proper positional data. First app is forest digital paintball.
I can’t wait to get my hands on this !!!!! This is so sick
6:20 always prefer bad English over low intelligence. Thank you! Very enjoyable watch
Great comment! Thank you!
@@Skyentific Just so you know, it's very easy to understand you and IMO your command of English is excellent. Great talk.
Thank you for interesting video! Cool!
Thank you for watching and for comment!
What cloud service did you use to retrain the model? Thx
What a very nice piece of hardware 😍
excellent stuff!
I'm really surprised by the amount of memory these things have. If they can cram 64GB @205GB/s in such a small module, or 16GB at 102GB/s on a SO-DIMM stick, why don't they do this with their desktop GPUs as well? You can never have too much video memory :)
Also, if you think about it, the Jetson AGX Orin are not too different from the Apple M1 Mac Mini.
You just have to pay the price for a RTX8000 or HM100...
Compared to M1, the Cpu isn't that great.
the 3090TI has a memory bandwidth over 1 Tb/s iirc at 24Gb or gddr6 so its probably just more power to make it faster and more of a bus width needed
@@AltMarc I know; I use the RTX 8000s at JURECA-DC for my work.
Still their ~$7k price tag seems way off. The memory chips cost a lot less. Hardware vendors could put much more memory on their GPUs and still be profitable. Right now, compute power rising much faster than memory capacity/bandwidth creates big issues for software.
The benefit of more/faster memory for consumers, video producers and especially scientists would be huge.
Bare in mind that the nvidia GPU cards are way more powerful than this embedded orin device, and consume up to 8x more power too.
I'm jealous
Same haha!
I have a 2GB Jetson nano only, the powerful ones were all sold out
The good thing that all Jetson have similar software. So if something work on one, it will work on another.
@@Skyentific true
Thanks!!!
Build a humanoid robot in the style of Aneca. The Jetson will fit nicely in it !
Where did y buy it? Could y send the link? In NVIDIA web sitevsays out of stock
It was sponsored by NVIDIA.
Sir, thanks for your share. I've got the same kit and set up the kit with NVidia SDK Manager (from host pc). System works fine, CUDA 11.4 folder is exist in usr/local but CUDA doesn't work even with my basic Opencv-Computer Vision projects. Is that normal ? So, am I suppossed to set up CUDA manually or with host PC
can you lease share the notebook you used for the first part with deep stream. thank you.
The jetson agx Orin, Can you test stable diffusion. And also a good video editing program like di Vinci. The ai suggest that you can run both on there. I have an old laptop and where as I’m not concerned with playing high end games it looks like this may be a good purchase to consider LLM running locally, stable diffusion, and general video editing 4k video for example.
I am working on a robot platform from scratch and god damn the moment i have a slight excuse to buy something like this i am pulling the trigger but i am not quite there
The speech recognition demo is quite cool, but it would be good if you test it in a noisy (outdoor?) environment to check if the model is robust and usable in real case scenarios really. I like when you conclude by saying "so far it works great" and it recognized "so fireworks..." haha, those models always screw up during those type of moments.
This is good point. I will test it.
2K USD? That's VERY expensive!
Agreed, though I guess that many companies who want to deploy the less expensive Jetson boards might still buy the Orin units for their developers to use at their desks.
It is expensive. But if you normalise the price to the performance, it is quite cheap (cheaper than Jetson Xavier NX for example).
can i connect to the jetson actuators and servos too?
Very interesting thank you. Depth and rgb camera should be included. I am sure it is comming as well as an extensive AI model to recognize common objects and escenarios like animals, trees and stairs. Finally it will be a general purpoise brain. And cheaper I hope jeje. Thank you
How to do the helmet training in cloud? Thx
This is great! I can finally program my NVR cameras to record humans and not kangaroos lol!
Really amazing but price is huge (
What a beast
Hi Skyentific, Great video!
question, is the Orin case made of metal or plastic mold ??
Can you use it to run text editing software on it like Microsoft Word?
Yes, you can run OpenOffice easily. :)
I would be curious to see it being used in transfer learning projects, where models based on RGB input are retrained to use thermal IR cameras, e.g. to track and recognise faces in the dark.
Or use it in sensor fusion between lasers scanners and (stereo) cameras.
Really depends on the model but since transformers are memory hungry, but this comes in 64GB shared vram memory (as far as I understood it), it should not be a problem as long you be mindful to not over max the available tensor (i.e. 4K video + multiple IR sensors etc). Thats inference. The second part regarding training I don’t understand - training a transformer model at the edge (something that is usually done distributed along many many machines, also because you need tons of training data) is not what I’d say this is good for. Aka Inference with large models: Yes. Training: No.
@@dinoscheidt I didn't mean doing the training on it, just using the new model in a new mechatronics project.
@@EatRawGarlic Yeah was confused as said. Anyway: You should try it and make a video about it. I’m curious too 🤓
you use existing models. it would be nice to create and train an own model, for example to detect cupcake and donuts. here you need a lot of computing power ... and if you can do this, you can delop unique applications. maybe it's nonsense what I'm saying here, in any case, many thanks
What you saying is not a nonsense. But it is complicated to make and train the model from zero. First you would need a lot of data, super powerful computer and knowledge of deep learning. That is why TAO toolkit is great. It allows to reuse existing models relatively easy.
What a cool video about something that is already out of stock and I will probably never be able to find or buy
Doing all this within 15w to 40w of power is truly amazing. My computer is much faster, but it consumes up to 1600watts of power. Not something you want to throw into a robot unless the robot is on a solar-powered golf-cart with a generator in the back seat, as a backup. Having all these "tuned" and "specialized" AI and CV components is the biggest advantage. Even if it hits 40watts, in the long-term it is consuming only a fraction of the same power as a typical CPU setup would consume, to process the same data. Doing it faster is a real nice bonus to the package.
It will not be long before we see a Jetson with the processing power of a 1080 or 2080 GPU, wrapped up in a nice 60watt package, for under $1000 in a kit, I am sure.
How much does it cost?
jetson nana
when you finally unbox your new most expensive jetson: th-cam.com/video/WRHIz97L2iI/w-d-xo.html
I have a RTX 3060TI which has much more power and I bought it for 480$. What's the point of 2000$ for Jetson AGX Orin?
Great question. First, for your RTX you need to add motherboard, memory, and CPU, this will be another 700$ at least (12core CPU+32G of memory). So your total cost will be more than 1k$. And your system will be way more bulky (Jetson module is only 100x87mm). Also, your PC will consume around 500W of power. For mobile robot this is a lot (only 1hour on the standard bicycle battery), Jetson consume only 60W at maximum, so it can run 8hours on a standard bicycle battery. Plus Jetson has some additional features: special port for CSI cameras (it is better than USB camera, as it requires less CPU to run); GPIO pins for sensors and other hardware.
Also there is at least one practical point: on mobile robots you can have some vibrations (due to the moving), and PCI-E slot does not works properly with vibrations. I know a company, which had a lot of trouble with this.
@@Skyentific "PCI-E slot does not works properly with vibrations" -- a very interesting datapoint, thank you. I wonder if you can share any additional info about this failure mode.
memoria eMMC instead of UFS. What i say to expensive.
Martinez Jason Rodriguez James Harris Charles
Rodriguez Jennifer Jackson Barbara Jones Helen
This thing is _$2000_ and is the same size as an ITX PC, and has very limited software support and ecosystem compared to regular compute platforms, so lots of fiddling with compilers and large chance this just kills your project. I kind of don't see the point of this thing, it's compute per $ isn't even competitive with a regular ITX PC with a mid-range GPU from 5 years ago, so what is the point? Maybe it's slightly less power hungry?
2400€? hahahha To expensive for what it give.
With normal PC can do more than that developed kit. and is only 4-5 times biger but i sure i dont have problems with the heat disipation.
🦾🦾🤖🤖🦿🦿🍖🍖😘😘