Is anyone doing deep learning on the NVIDIA Jetson TX2?

Discussion in 'Machine Learning, Deep Learning, and AI' started by Patrick, May 11, 2017.

  1. Patrick

    Patrick Administrator
    Staff Member

    Joined:
    Dec 21, 2010
    Messages:
    11,560
    Likes Received:
    4,490
    Just wanted to see what folks have been using successfully on here with the NVIDIA Jetson TX2. (e.g. is there a good Docker image to get me started?)

    Also, let me know if anyone is interested in content/ video on the board and such. I would be happy to do some content for the main site with these if folks are interested.

    I know it is not smart to train models on the Jetson TX2 but I want to see what it looks like doing so.
     
    #1
  2. MiniKnight

    MiniKnight Well-Known Member

    Joined:
    Mar 30, 2012
    Messages:
    2,951
    Likes Received:
    860
    Siiiick! That's such a huge upgrade over the TX1. Newer CPU, GPU, 2x RAM capacity and BW, 2x storage.

    I'd like to know a few features

    - Can you use a NVMe drive?
    - Can you use a 10gb NIC?
    - How's the software stack?
    - I'd love to know how it works as a thin client too.
    - I'd want to know if you can do FFmpeg / libav with it or do a Linux GPU accelerated desktop.

    I don't get why you'd train on this but if it's just to get benchmark numbers fine. It's only 1/2.5 of a 1050 Ti. Low power's nice but not if you've gotta scale training.

    Inference great tho. I get that. Hook it up to a camera and do recognition.
     
    #2
  3. TangoWhiskey9

    TangoWhiskey9 Active Member

    Joined:
    Jun 28, 2013
    Messages:
    402
    Likes Received:
    59
    I'd like to know power consumption, if they can fit in a 1U short depth case, if they support virtualization and whether they can be used for standard Linux apps.

    I know it sounds psycho. 8GB RAM and 64GB onboard eMMC and it's a cheap way to run a directory server when you're not using the GPU. If you can hook it up to storage and do transcoding, that's great.

    On the ML and AI side I'd like to see TensorRT and Vision works VisionWorks

    They come with the dev kit NVIDIA Jetson TX2 Delivers Twice the Intelligence to the Edge | Parallel Forall

    DOES IT DOCKER?!?!
     
    #3
  4. LukeP

    LukeP Member

    Joined:
    Feb 12, 2017
    Messages:
    161
    Likes Received:
    15
    All gpus are the same, limited by sgemm. Nothing new in AI for last 7 or so years really. Just faster lol
     
    #4
  5. Patrick

    Patrick Administrator
    Staff Member

    Joined:
    Dec 21, 2010
    Messages:
    11,560
    Likes Received:
    4,490
    The one part I would disagree there is that there is a huge difference between AMD and NVIDIA GPUs. NVIDIA = relatively easy to get going with just about every guide out there. Tensorflow default will use NVIDIA and AMD support is not there. mxnet supports NVIDIA easily and AMD support is not there. I offered to pitch in money for Raja at AMD to do a Kickstarter for amd-docker a few months ago.

    A hyper-scaler and big AI research team infrastructure guy gave me this hierarchy:
    • NVIDIA = top of the world right now
    • Intel = making a good effort and has some interesting tech
    • AMD = looks like they are going to miss the AI train if they have not already missed it
     
    #5
  6. LukeP

    LukeP Member

    Joined:
    Feb 12, 2017
    Messages:
    161
    Likes Received:
    15
    Up until just recently where we got FP16 and Volta's special tensor units, AMD's hardware was better than Nvidias on an aggregate level. eg comparing an 1000 gpu AMD farm to the equivalient $ gpus of Nvidia.

    But its not all about the hardware thats for sure. There is a huge difference in software ease of use, you are right Patrick. The Tools have developed over the years. I use my own software tools so I tend to only notice the hardware changes!
     
    #6
  7. Patrick

    Patrick Administrator
    Staff Member

    Joined:
    Dec 21, 2010
    Messages:
    11,560
    Likes Received:
    4,490
    Some key bits on the Jetson TX2:

    1. Docker installing is broken at the moment it seems.
    2. DO NOT EVER sudo apt-get upgrade -y if you do, you need to do a full recovery.
    3. If you do a full recovery, you will need to wire your own serial interface plus connect the Jetson TX2 via USB to do the OS load
    4. You apparently need a Ubuntu desktop to upgrade. I would not recommend using a server unless you are doing USB pass-through as well. Even then, it is a bit scary.
    5. Once you do get the Ubuntu desktop hooked up to the Jetson TX2 it works well
    6. Serial cable to see what was going on required a trip to the store to get one that would accept the custom pin out of the Jetson TX2. Of course, the cheap one I could purchase on the spot had different wiring colors than the guides online.
     
    #7
    Abhishek Kumar likes this.

Share This Page