Author Topic: Jetson nano 2GB for $55, or 4GB for $99, or something else?  (Read 1155 times)

0 Members and 1 Guest are viewing this topic.

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Jetson nano 2GB for $55, or 4GB for $99, or something else?
« on: February 17, 2021, 02:02:48 pm »
nVidia Jetson nano SDK has 2 flavors, same processor, the main differences are:
   - $55 for 2 GB RAM, 1 x USB 3, 2 x USB 2, 1 x MIPI, 0 x M.2 key E
   - $99 for 4 GB RAM, 4 x USB 3, 0 x USB 2, 2 x MIPI, 1 x M.2 key E

https://www.seeedstudio.com/blog/2020/10/05/compare-nvidia-jetson-nano-2gb-and-jetson-nano-4gb/

The idea is to fool around with vision and machine learning, and don't know how big of a difference will make in practice a 2 vs 4 GB of RAM for the overall performance.  No hands on experience, and most probably the board will end up in a drawer after a few weeks, so I'm not sure which one to buy.

Anybody who worked with Jetson nano, what to buy, please, the Jetson nano with 2 GB, with 4 GB, or maybe something else than a Jetson nano?

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #1 on: February 17, 2021, 07:53:12 pm »
Just for the records, it looks like the nVidia Jetson nano has the most powerful image processing in its price range.

However, for the Jetson nano SDK $99/4GB version, I've found that adding an M.2 Wi-Fi card and a decent speed uSD card will make this buy the same price as a cheap new laptop, and the laptop will have way more hardware and the same if not faster GPU than a Jetson nano has.

The $55 (or $59 with an USB WiFi dongle included) Jetson nano 2GB would make sense for the price, especially for an autonomous vision-based robot or alike, because the SDK module is small and light enough to be carried by a moving robot.

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #2 on: March 07, 2021, 01:59:56 am »
I don't know anything about the Jetson but I have an interest in AI so I bought the 4GB version along with the WiFi attachment and a case.  The metal case is from Seeedstudio, everything else is from Amazon and will be here in a day or so.

I have been especially interested in CUDA programming and parallel Fortran.  I'm not sure that all of that is actually working so I thought I would start with the Jetson and its SDKs.  We'll see!

There are a lot of reference projects on the nVidia site and some others at Seeedstudio.  There should be enough to get started.

 

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #3 on: March 07, 2021, 09:56:37 am »
Good choice to get lots of AI fun!   :-+

Good choice for going for the 4GB version instead of the 2GB.  I've seen in some of the nVidia tutorials they were manually increasing the swap file/partition for the 2GB Jetson, which makes me think the extra 2GB on the 4GB Jetson Nano will not just sit there unused.

Not sure how much can a Jetson help with Fortran, AFAIK CUDA is advantageous for matrix and multiple data single instruction type of computation, but I know nothing about the subject, so don't mind me.

If you are into crunching numbers, I think a recent nVidia video card will be faster, but also much more expensive.  However, even if a Jetson module can not beat the latest 1k$+ GPU, it can still be a ton of fun and learning, especially with robotics and vision.  A Jestson is small and low power enough to sit on a mobile robotic platform, yet more than enough to do offline inference and even offline NN training.  Robots with vision are cool.

One of my forever postponed projects is a gesture to SCPI bridge (for example to operate the oscilloscope by making visual gestures).  That would be almost trivial to do with a webcam and a Jetson, and more important, it would be good enough project to justify my impulse buy.   ;D

A second forever postponed project of mine (vision based) is an universal data logger, a webcam that watches the display of a DMM or some other instrument that needs to be logged.  While researching the Jetson nano, I've learned inference can be done even on very modest hardware, for example somebody made a visual data logger using a $10 (including the camera) ESP32-CAM module.   :o
https://hackaday.com/2021/02/07/an-esp-will-read-your-meter-for-you/
https://github.com/jomjol/AI-on-the-edge-device



Back to Jetson nano, it's nice that nVidia made containers with preinstalled AI toolchains, that spares a lot of time.  Their website seems pretty active, too.  There is also a place where people can post their Jetson projects, a lot of videos and tinkering ideas.

https://developer.nvidia.com
https://forums.developer.nvidia.com

You may want to look at their "nVidia Developer" Youtube Channel, too, they have a few playlists about robotics and applied AI with Jetson (mostly about vision and self driving), e.g.:

https://www.youtube.com/playlist?list=PL5B692fm6--ufBviUGK3hlwL1hVSyorZx



Also, there are some free classes from nVidia if you register on their website to enroll:
https://courses.nvidia.com
« Last Edit: March 07, 2021, 10:36:56 am by RoGeorge »
 

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #4 on: March 07, 2021, 04:01:55 pm »
I was playing with the Fortran DO CONCURRENT statement the other day.  The compiler unrolls the DO loop and the scheduler dispatches threads or tasks to perform each iteration.  In a trivial example, 'top' shows 10 tasks open and close when executing something like

DO CONCURRENT I=1,10
  <some calculation>
END DO

nVidia extends (or implements) the idea and it is possible to execute the loop on CUDA units:
https://developer.nvidia.com/blog/accelerating-fortran-do-concurrent-with-gpus-and-the-nvidia-hpc-sdk/  Even the Jetson Nano has 128 CUDA cores available for execution (the count is right, I don't really know how it works, yet).

The cool part is that the dispatcher will send fragments anywhere.  It can be another task on the same computer, CUDA units on the same computer or even distributed computers sort of like SETI at Home.

Fortran is the only standardized language with built-in support for concurrent programming.  It's also what I started with back in '70.  Maybe there is some way to use it with JetBot.

I have been thinking about buying an upgraded video card but that seems a little excessive at the moment.  First I would have to think about upgrading the system itself.  The PC was fast a few years ago, there are better available today.  Maybe after I play with the Jetson...

Then there is Python...  It seems all the cool libraries for AI are targeted at Python.  Not my favorite language but I'll manage.
 

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #5 on: March 07, 2021, 05:42:11 pm »
AFAIK many (if not most) of the Python ML and AI libs/engines/frameworks are in fact written in C/C++, but usually it's their Python wrappers who get the most fame with demos and tutorials, because C will scare away the newcomers.

Same with numpy and matplotlib, they are just wrappers over C libs, many of which are coming from the Fortran's numerical algorithms.   ;D

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #6 on: March 09, 2021, 04:38:51 pm »
Initial boot worked out ok.  I had a wired keyboard and the system accepted my wireless mouse.  I had an Edimax SiFi dongle and it worked fine.  Don't even think about plugging the dongle in after boot.  I had a horrible time trying to get it to register with the router.  I rebooted and it all worked fine.

I am using a 256 GB microSD card.  I haven't decided whether to leave it that way or change over when the 64 GB cards come in later today.

The system image comes minimized and there is an option during setup to add in the packages that were considered redundant to needs.  I did that since I had a LOT of drive space.

Too bad that nVidia chose Ubuntu with the Unity desk top.  I absolutely hate that distribution where the system buttons are on the wrong side of the window.  In the old days, you could do something with a file in your home directory to put the buttons back on the right corner.  I understand that feature has been removed.  I need to research this a little more.  Meanwhile, I don't have anything good to say about the arrogance of the developers.  It was bad enough when I had to change the location and now I can't even do that!

I haven't gotten far enough to know how much time is spent with the desktop.  The Jupyter notebooks run out of a browser so the display may not even need to be on the same system.  Just web browse to  <ip>:8888

Meanwhile, it is nice that the system comes right up when I followed the directions.

The board can be powered by USB over the microUSB cable carrying the serial stream.  However, it will soon throttle due to excessive current.  I plugged in a 4A 5V supply and the warning was cancelled.  The jumper at J48 is pre-installed.
« Last Edit: March 09, 2021, 04:40:29 pm by rstofer »
 

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #7 on: March 10, 2021, 04:10:10 am »
Too bad that nVidia chose Ubuntu with the Unity desk top.  I absolutely hate that distribution where the system buttons are on the wrong side of the window.  In the old days, you could do something with a file in your home directory to put the buttons back on the right corner.  I understand that feature has been removed.  I need to research this a little more.

AFAIK it should be possible to move the buttons from left to right from the theme, but in the last couple of years I only used Kubuntu, which is KDE Plasma and has way more many settings.  Ubuntu should have some sort of gtk based GUI, I remember there were a few GUI tools for advanced tuning, thought it might still need file editing tinkering, for example to make the scroll bar thicker I had to change some CSS files back then, like it would be a webpage but it's the desktop.

Since the SD card is big enough, try to install KDE Plasma for one of the heaviest, or LXDE as a lightweight desktop GUI.  There can be more than one installed all together.  Since that is an ARM based PC, there might be less binary tools in the repositories than it would be for an x86, but you can always compile from sources.

Alternatively you can just download and install on another card the OS image for Jetson nano 2GB.  That one is the same, except it has less RAM, so they used LXDE (Lubuntu) instead of GTK (Ubuntu).

Congrats for the new toy!   :-+
« Last Edit: March 10, 2021, 04:14:06 am by RoGeorge »
 

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #8 on: March 10, 2021, 06:06:57 pm »
I misstated the J48 jumper; it is not actually connected to both pins, it is dangling off one pin.  The CPU will almost immediately go into overcurrent when fed from the microUSB and doing much of anything.  Installing the jumper properly and using a 4A supply solved that problem.  Or I could run it with just 2 cores - like that's going to happen!

I finally got all the gadgets to begin work.  We'll see...  It looks pretty nice on a 27" monitor!
 

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #9 on: March 11, 2021, 04:11:02 pm »
There is something wrong with the distribution image.  It works just fine but it doesn't include the tools like Jupyter which seems to be what the demo videos are using.  You can chase this forever across the various forums without resolution.  One approach is to download the SDK and build an image.  Problem is, when it comes time to select the version of Linux to build to, there is no choice in the selection box.  The instructions mention connecting the Nano to the host computer but it doesn't say how.  Over the serial port?  It's already connected over Ethernet but this doesn't seem to be what they want.

The SD Card Nano image file (tagged "For Jetson Nano Developer Kit:") is the one that is broken:
https://developer.nvidia.com/embedded/downloads

The SDK approach is a work in progress - very little progress, lots of work.

There are discussions all over the Internet without resolution.  There is one that mentions you have to use the DLI image that comes with the DLI course which comes with a $90 price tag.  Maybe the proper image is in there, maybe not.

There are threads where it is recommended to avoid installing the tools separately.  Known directory locations being just one of the problems.

More to follow...  If I come up with an image or solution, I'll post back.
 
The following users thanked this post: Michael Rempel

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #10 on: March 11, 2021, 08:01:10 pm »
OTOH, the CUDA examples in the basic image all work.  It took a little over an hour to compile the bunch of them.

Here's a video of the initial setup and some CUDA demos. 

https://youtu.be/BkZ1n_1F-Cg

Here's the matching article with the CUDA examples build instructions about 2/3 the way down.  The oceanFFT example is terrific but the smokeParticles has to be the most impressive due to the ray tracing and the shadows on the floor.  Amazing!  The Mandlebrot and Julia Set examples are also interesting.

https://dronebotworkshop.com/nvidia-jetson-developer-kit/

Here's a metric:

Control Data 7600 (the biggest and baddest super computer of the early '70s had throughput of about 10 MFLOPS and rented for around $100k per month.  The Jetson Nano (the entry level device) has throughput of about 473 GFLOPS (47,300 times faster) and costs $99.  Sure, it's missing the rows of tape drives and washing machine size disk drives  but with a 256 GB microSD and a couple of USB thumb drives, it has equivalent storage.

47,000 times faster and less than an hour's rent.

The CDC 6400 had much lower performance (about 3 MFLOPS) and got us to the Moon and back.  Well, to be fair, there was more than 1 machine.  The Nano is 157,666 times as fast!

 

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #11 on: March 11, 2021, 08:30:34 pm »
There is one that mentions you have to use the DLI image that comes with the DLI course which comes with a $90 price tag.  Maybe the proper image is in there, maybe not.

No.  Jupyter lab is free.  It can be installed in many ways, all free.  The easyest would be with conda, or Anaconda, IDK what nvidia is using, I don't have a Jetson.

Not sure what are you calling SDK, and if you are trying to compile or to cross-compile.  Do you have conda or Anaconda already?

But anyway, Jupyter server should be already installed and running on the base install, IIRC, not inside the container.  Nvidia was saying they prepared various Docker containers with all the tools, maybe you are starting the Docker container without the proper parameters, IDK.  I have no idea how to help or what you are trying to start, and where from, and what exactly doesn't work and with what error.  Too few details.  Maybe the Jupyter server has crashed, maybe it's not started, or maybe the container doesn't have access to it.

Is the jupyter server (you want to connect the notebook to) supposed to be runing inside the container, or outside, on the OS?

Does the same lesson and its Notebook run if you use the original OS SD card image (with the OS image for the 4GB module)?

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #12 on: March 11, 2021, 10:42:29 pm »
No.  Jupyter lab is free.  It can be installed in many ways, all free.  The easyest would be with conda, or Anaconda, IDK what nvidia is using, I don't have a Jetson.
Yes, that would be the way to install it and the default image does have conda but in the various forums, this type of installation hasn't been shown to work out due to specific directories not being created but required for the system.
Quote
Not sure what are you calling SDK, and if you are trying to compile or to cross-compile.  Do you have conda or Anaconda already?
When you look at the download page there are 3 buttons.  Two are for the default image, 4GB versus 2GB over on the left and, over on the right, is a button for the SDK which ultimately builds an entire image with cross-compilation.  It also requires Ubuntu 18-4 for the host and I don't know where it gets the source for creating the image.  It will take a lot more research.
Quote
But anyway, Jupyter server should be already installed and running on the base install, IIRC, not inside the container. 
Jupyter server is not installed in the base image.  I'm not the only one complaining about it so I am guessing that it's not my screw-up.  So far I haven't found the PhD instructions (Push here, Dummy).
Quote
Is the jupyter server (you want to connect the notebook to) supposed to be runing inside the container, or outside, on the OS?
I have no idea.  That's going to be part of the research.  Somebody, somewhere, must have this stuff running.  I'll find it sooner or later.
Quote
Does the same lesson and its Notebook run if you use the original OS SD card image (with the OS image for the 4GB module)?
Not as far as I can tell.  It's a work in progress that got sidetracked with the CUDA examples and now the Sparkfun robot kit.  I'll get back to the Jupyter stuff at some point.  I very much want to follow along with the video series (the free ones).

Actually, once I convince myself that I can understand what is going on with the free examples, I may pay the money for the DLI course.  Education has never been free and $90 isn't all that much.
 

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #13 on: March 12, 2021, 10:19:04 am »
Education has never been free

Education has always been free.   :-//
(where I live, during my lifespan and I'm grateful for that - I was very lucky).

Just saying, not trying to go into any ideological debate about how it should or it should not be.    ;D

From school to PhD and postdoctoral, there are even scholarships for the very good students.  In theory this could be abused by freeloaders to stay in school forever, in practice it is not that easy, adulthood has its own rules and demands.

Of course, after the fall of the Iron Curtain in 1989, private universities appeared in Romania, too, and now there are payed classes, too, not only the free ones.  But here it is quite different, and many still sees the payed study as a second class diploma, because those who don't pass the admission exams for the free universities usually does not give up, and go for a payed university, where the entry bar is lower, or sometimes there is no entry requirements other than having a high school diploma.

The end result is that nowadays an employer will rather prefer to have employees coming from free universities.  Nobody is saying that out loud, but the preference is obvious.

Anyway, since I know nothing about NN (Neural Networks), ML (Machine Learning) and AI (Artificially Intelligence), I feel entitled to share my ignorance.   ;D

From what I've seen (by just keeping an eye on the field over the years, later watching free youtube classes, etc.)
- it was a big fizz about NN 50 years ago, or so
- then somebody managed to convince everybody else that NN will never work as hoped (I think it was something related with the XOR function, but don't blame me if it was not that)
- the pessimism spread and the field was abandoned for a few decades (the so called "AI winter" era)
- then everybody realized that proof was wrong, in the sense that it doesn't really matter
- meanwhile the hardware was making huge progresses in terms of computing power and available memory, and suddenly everybody was playing with NN and AI again

- the joy was big, hopes were yet again higher than never, research money start to pour in the AI field again, DARPA $1 million+ range prizes competitions for unis and individuals, etc.
- there was only one thing still missing for the AI to fully unleash, and that was the lack of data.  A NN is useless without training, just like a computer is useless without software.  There are many ways of AI training, but all of them boils down to exposing the NN to piles and piles of data, hundreds to billions of examples to teach the NN what to do and what not to do

- and here it comes the need for data harvesting.  Again, the times are just right for the data harvesting frenzy we are living in right now.  Ubiquitous Internet access, a computer in any pocket, hive connectivity for the masses.

- the moment seems just right technologically, decent computing power, cheap hardware, big data

We were hunters and gatherers tribes, then agriculturists and salesmans, than craftworkers and mechanicists, then electronists and programmers.  Guess now we are heading to the AI era.

 :=\

The future is now!
So where's my flying car!?   ;D




Oh, and all the NN is about fitting curves through the data while learning, then just crunching numbers during inference, when the function of that curve we just fit is applied to each dataset in order to compute the outcome.

Training is the hard thing to do (in a reasonable amount of time), same as learning with humans.  This is where most of the effort is now, and IMHO there will be no magic trick here, just hard work.

The big hope is that once a NN training is done, we know how to replicate that learning as cheap as would be to copy a file, and suddenly we can have a flood of top notch expertise available to everyone, like we have now the flood of data since "the Internet".   8)

Offline DiTBho

  • Frequent Contributor
  • **
  • Posts: 730
  • Country: gb
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #14 on: March 12, 2021, 12:13:32 pm »
Education has never been free
Education has always been free.   :-//

That's life. My educational cost 50.000 Euro for the university and external courses.

 

Offline DiTBho

  • Frequent Contributor
  • **
  • Posts: 730
  • Country: gb
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #15 on: March 12, 2021, 12:24:49 pm »
At the moment I have been working on a project that has two PowerPC-40x boards running at 200Mhz and only 256Mbyte of SDRAM. No FPU, no GPU, but a couple of free PCI32-5V slots where I am planning to install a PCI-FPGA card for offloading some computations.

That's the version1 of this secret project, and it's made with 90s and 2000s technology, but it's not secret that I am going to use 2020s technology for the version2.


Probably two Jetson Nano in parallel, plus other stuff.
 

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 3073
  • Country: ro
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #16 on: March 12, 2021, 03:34:44 pm »
Probably two Jetson Nano in parallel, plus other stuff.

If it were to need more than one Jetson, maybe the other architecture of nVidia SBCs would be more suitable, called Xavier, based on Volta architecture: 
- $400 for Xavier NX w 6 ARM cores, 8GB RAM,  NVIDIA Volta, 384 CUDA cores, 48 Tensor cores
- $700 for Xavier AGX w 8 ARM cores, 32GB RAM, NVIDIA Volta, 512 CUDA cores, 64 Tensor cores

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #17 on: March 12, 2021, 03:57:04 pm »
Education has never been free
Education has always been free.   :-//

That's life. My educational cost 50.000 Euro for the university and external courses.

And books and computers and, most precious of all, time.  There's only so much time and a lot to learn.

The GI Bill paid for my undergrad and my employer paid for grad school.  In that sense, my education was 'free' but there is the matter of 5-1/2 years of time.  Sure, I recovered some money in wages for the time spent in school but I had to make the investment up front.

« Last Edit: March 12, 2021, 04:00:14 pm by rstofer »
 
The following users thanked this post: DiTBho

Offline DiTBho

  • Frequent Contributor
  • **
  • Posts: 730
  • Country: gb
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #18 on: March 12, 2021, 04:45:32 pm »
If it were to need more than one Jetson, maybe the other architecture of nVidia SBCs would be more suitable, called Xavier, based on Volta architecture: 
- $400 for Xavier NX w 6 ARM cores, 8GB RAM,  NVIDIA Volta, 384 CUDA cores, 48 Tensor cores
- $700 for Xavier AGX w 8 ARM cores, 32GB RAM, NVIDIA Volta, 512 CUDA cores, 64 Tensor cores

Thanks. I will consider it  :D
 

Offline rstofer

  • Super Contributor
  • ***
  • Posts: 8309
  • Country: us
Re: Jetson nano 2GB for $55, or 4GB for $99, or something else?
« Reply #19 on: March 12, 2021, 09:50:46 pm »
A more complete image file is at https://developer.download.nvidia.com/training/nano/ainano_v1-1-1_20GB_200203B.zip

username: dlinano
password: dlinano

It is a 20GB .zip file versus the default 6GB file so there is bound to be more stuff loaded.  Indeed there is!  Among other things, the Jupyter server is configured and the swap file is set to 6GB versus none by default.  I'm using a 256GB microSD so I really don't care how big the image is.

Now maybe I can play along with the free tutorials.
« Last Edit: March 12, 2021, 09:54:01 pm by rstofer »
 
The following users thanked this post: Michael Rempel, DiTBho


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf