Everybody And welcome to a review slash overview of the new end video Jetson Nano death kit.

Now, historically, these deaf kids have been on them or expensive side, like 200 to $400 or even more than that.

But the man who comes in at $99 which in my opinion, approaches MME.

Or of the general hobbyist or the type of person who might be buying raspberry pies instead.

And what you're getting for this $99 is pretty darn awesome.

So, first of all, what do we get it?

So the biggest thing that we're getting here over say a raspberry pi is a 128 core Maxwell G p You also you're getting a quad core arm.

CPU.

You've got four gigabytes of Ram Micro SD storage.

You've got four USB 3.0 ports.

You've got tons of G p I O.

Here.

You've got a C.

S.

I connector here for cameras.

So something like the raspberry pi camera also, you could just use you as be cameras.

Of course.

Ah, we've got gigabit Internets.

We've got HTM I display ports.

You've got a five volt for amp in here.

You can also actually power this board via micro USB.

Can't imagine why you'd buy this board for anything that would allow you to use that low of power.

But, hey, it's there for you if you have that use.

So first of all, you know, this thing is called Nano, so I feel like I probably need to address the size of the board.

So I have here not a raspberry pi for, but it's same size.

It's a raspberry pi three, so you can kind of see how these compare in size.

Obviously, due to this heat sink, it's actually considerably and considerably taller, although most people do wind up putting something on their, um, their processor there.

But, uh, you actually don't have to with a pie.

I'm gonna talk about that in a second here with the Nano, Um, so it's considerably taller and it's about me.

It's less than double the size of a pie, but it's It's close to double the size of the pine, but where it gets more interesting is comparing it to sizes of the old board.

So, for example, this is a T.

K one so really old dead kit board.

But you can see here it's considerably smaller than even this board, which was actually one of the smaller ones.

And then in comparison, I don't have one.

But the T X one in the TX two of just drawn out here.

So here's our little baby Nano and then you've got your T K one and then the T X one in two sizes for a comparison so quite small can definitely be used in a lot of applications because of it.

So, uh, let's talk briefly about heat.

So traditionally with raspberry pies, I've kind of abused them.

I don't really cool down the processor in any way.

I never put fans.

Well, I've seen people with little heat sinks.

It's cute.

It's adorable.

It's not necessary of used raspberry pies and very hot blistering conditions.

They are really robust boards.

They're totally fine.

But, um well, I mean, okay, I have killed my fair share of raspberry PiS.

Don't get me wrong, but the raspberry pi is ah, very hardy board.

Now, when it comes to the Nano, I don't know if it's just because the GPU runs hotter or what?

Even including this gigantic heat sink.

You still have to have a fan.

You just do it at least if you're using that GPU, which is kind of on because historically, all of the deaf kits I This is only when I have, but they've all come with fans.

I'll put up some pictures if I remember two of the T X one and t X two, but they have heat sinks.

But then there's like a recessed fan in there.

I really wish they would have just done that to this board.

I don't know if it's a cost savings thing or what, but I strongly advise that you get some sort of extra cooling.

So I'm thinking maybe some sort of water based solution you?

No, no, I'm just kidding, but definitely some sort of bigger heat sink or something like that.

Okay, but seriously, you definitely want one of you's like little fans or something like that that you can put on, or a case that has a fan.

You'd need to be moving air.

If this thing is just sitting here, this heat sink heats up.

It's like burning to the touch and within about I would say it took about 15 minutes.

I was running an object detection model, and in about 15 minutes, I front the board just like, turned off the SD card got corrupted, and I had to, like, re mount the image and all that.

It was It was really fun.

But this thing was super hot, like using a really advanced temperature sensor called my thumb.

It was too hot to the touch, which is generally a bad thing for especially when it's the heat sink that is hot to the touch causes distributing, um, heat over this entire heat sink, which I promise you is not the size of your CPU or your J P.

You.

So yeah, um um big deal.

Be really careful about the heat and just buy a fan or use a case with a fan or whatever.

All right, so now what I want to do is just show you guys a quick example of doing object detection on this board were quite literally going to grab tensorflow.

Put it on here, run the tensorflow object detection a p I and see what we get.

And then the other thing I'd like to do is show you guys one of the newer technologies from and video Cold tents are our tea.

And, uh, when we get there, I'll explain a little bit more about it, so let's check it out.

All right?

As silly as I feel recording physically the monitor.

What we're looking to get here is an actual frame rate for the object detection model.

And if I was using something like O.

B.

S to record the screen, that would be skewing results considerably.

So what you see here is straight tensorflow GPU We're running the tensorflow object detection A p I, and we're using the inception V to model atyou can see here.

We're getting 1.3 frames per second with one object detected.

The more objects that you detect that's going to impact frame rate.

So we definitely want to compare either with zero objects, which I think is kind of silly or one object at least or more.

So for here, we're just gonna compare with one object at a time.

Anyway, you can see 1.3 frames per second.

Now, most people would be like, whoa.

Okay, but shouldn't you be running TF light and not regular full bloom tensorflow on such a tiny device, and that's probably true.

So now let me go ahead and set up for T f light.

Okay.

And here we have the TF light example again, using wth e inception V to model.

And what we have here in this case is somewhere between four and 4.1 frames per second.

So a considerable improvement.

I mean, you can see I'm just moving this pen around just for example, of frame rate.

That's not terrible.

I should have done that on their one frames per second, but at one frame per second.

The problem is, um, basically your reactions and the things that you're seeing could be delayed by an entire second, which is pretty, pretty hefty.

So in this case, we're seeing four frames per second.

So not too bad.

That's borderline dare I say acceptable.

Now let's show the tense or or t version.

So here we have the tensor arty version of inception V two.

It's getting on average like 9.5 frames per second.

It's fluctuating from the upper eights to even 10 frames a second, so I'm gonna call it 9.5.

Um and that should be fair.

I think eso compared that to say, the, uh the TF light.

It's about twice as fast and then raw tensorflow GPU.

It's like nine times faster.

So that's pretty cool s o.

I'll just pushes petting back and forth so you can see, you know, I wouldn't want to play a game at this frames per second, but it's fairly reactive.

So again, this is all done using tense or or tea, which is an optimization that we can apply at the basically once we better frozen tensorflow graph, we can use tents are artie to further optimize and hopefully speed up significantly.

Inference which basically, that's when we're doing our predictions.

So it's not gonna help us speed up training or anything like that.

But in terms of actually using it in practice, it can be considerably faster than what you might just get via just your vanilla output on dot predicts so really, really cool stuff I hope to actually look into tents are arty, much more in the future s O, for example, with like the self driving car and grand theft auto, the more friends per second we could get the performance would increase like considerably like we go from 30 frames.

A second toe like 45 made a huge difference.

So seeing improvements like this really excites me.

So I don't want to look more in a tense or Artie into the future, but also pretty cool to see it helping out this little nano to object detection.

Finally doing object detection on the Nano.

I just chose Inception V two.

It's a decent model.

Feel free to try a different model if you want.

I definitely This is not the quickest model to run on a small little device like this, so you can definitely get way better frames for a second more.

You can get for sure, like 30 frames 40 frames plus, um, but this is an actually pretty accurate model that you're not gonna get either miss a lot of things or get the wrong classifications.

Stuff, questions, comments, concerns, whatever.

Feel free to leave them below.

Come join us in the discord.

That's discord dot gov slash Centex.

Otherwise, I will see you in another video.

This video is brought to you in part by the following Amazing Channel members ready In Edward's Mike Smith age a Ch'ti Santa knew Bo Mikhail Gun be the dragon.

573 Sir.

Sroka s all with us for a month now as well as Kunal.

Muggle con in Christiansen.

Visit Iran.

Kumar.

Why, you, Michelle Berardi with us for two months in finally drag a duck's Juma Khan sank more three months.

Thank you all very much for your support.

Helps me put out awesome videos like this for awesome people like you.