3689712360?profile=original

https://devblogs.nvidia.com/parallelforall/jetson-tx2-delivers-twice-intelligence-edge/

NVIDIA is pleased to announce Jetson TX2, the world's preeminent embedded computing platform for deploying deep learning, computer vision, and advanced artificial intelligence solutions to edge devices in the field. Jetson TX2 delivers server-grade performance that fits in the palm of your hand and runs on less than 7.5W of power.

Driven by integrated NVIDIA Pascal GPU with more than a TFLOP/s performance and hex-core CPU complex with dual-core NVIDIA Denver2, quad-core ARM Cortex-A57 and 8GB 128-bit LPDDR4, Jetson TX2 includes user-tunable energy profiles (Max-Q and Max-P) and is built from the ground-up for ultimate compute efficiency. Jetson TX2 is packed with hardware multimedia engines for streaming high-bandwidth data and 4K inputs/outputs, including up to six simultaneous MIPI CSI cameras, Image Service Processors (ISPs), video encoders and decoders supporting H.245/H.245/VP8/VP9, and additional high-speed I/O like PCIe and USB 3.0.

NVIDIA JetPack 3.0 AI SDK provides comprehensive software support out of the box, including CUDA Toolkit 8.0, cuDNN 5.1, TensorRT, Linux kernel 4.4, and Ubuntu 16.04 LTS. Jetson TX2 is the ideal platform for deploying deep learning frameworks like Caffe, Torch, TensorFlow, and others into an embedded environment.

Developers can get started deploying AI with NVIDIA Two Days to a Demo, a tutorial on GitHub which provides neural network models and deep learning vision primitives like image recognition, object detection, and segmentation, and teaches the workflow for re-training the models using NVIDIA DIGITS interactive training system. New to Two Days to a Demo for TX2 are segmentation models geared for drones and an aerial training dataset to encourage development of autonomous flight control systems. To learn more, see my in-depth technical Parallel Forall blog on the NVIDIA site.

3689712321?profile=original

Jetson TX2 is available as a developer kit and module. The module is $399 in 1K quantities, and the Jetson TX2 Developer Kit is $599, with the $299 Jetson Educational Discount for those belonging to academic institutions. In addition the price of the Jetson TX1 Developer Kit has been reduced to $499. Preorders are available with shipments in North America and Europe beginning March 14.

Key to the recent resurgence in artificial intelligence, it's been exciting times at NVIDIA experiencing the vast increases in compute horsepower put at the fingertips of developers everywhere. With Jetson comes the ability to deploy advanced deep learning capabilities into embedded environments, onboard remote edge nodes. We hope you'll join us to begin developing your own AI-powered smart devices and computer vision solutions. See our full blog post with the details to learn more.­­

You need to be a member of diydrones to add comments!

Join diydrones

Comments

  • Really hoping to upgrade my TK1 for some decent machine learning performance. TK1 was such a fun thing to learn with, but now i'm at a stage where I really need to get my feet wet with Tensorflow dev projects. Such a strain trying to hack & build from source Cuda 6.5/32bit/TK1 to work. Ooof.

    My TK1:
    http://dalybulge.blogspot.co.uk/2015/12/ghetto-dji-matrice-manifold...

  • The consumer Teal drone is built around Jetson module and shipping en masse this spring for $1399. It's capable of sustaining 70mph+, so ability for fast processing and collision avoidance is important. Yes, industrial inspection and delivery drones a natural fit too. BTW all the numbers from the article are official from NVIDIA perflab, checked and re-checked, so nothing alternative about that. I am also thinking about us offering the half-price edu discount promotion to DIYDrones like we did in MAKE Magazine...for now, contact me directly if you have the need.

    Homepage
  • @DustinFranklin While you've gone a little donald trump/alternative facts with your marketing I (genuinely) appreciate your enthusiasm for your product.  I'm putting an order in for a tx2 through the edu discount, which is a really great program.

    @JohnArneBirkeland Yes you're right, the tegra stuff has never really been targeted at the DIY/hacking crowd, more at the automative/industrial embedded market.  Hopefully the modules alone will be more widely available, and with the tx1 at a lower price and lower price carriers coming on market, they will be more accessible in the future.

    I suspect we'll see the tegra products going into large industrial drones, whereas the likes of the joule is far more suitable for smaller diy/hacker drones (and for larger industrial drones where massively parallel fp computation isn't at a premium).

  • 3D vision, situational and environmental awareness, SLAM and intelligent on the fly navigation in an environment full of unknown obstacles.
    Parallel GPU processing is the real time key to that and that is the future.

    We have already done absolute position in an unobstructed environment with GPS, but the real interesting stuff is down here in our real world.

    And that takes serious CPU horsepower and something that won't drain your battery in 5 minutes.

    The TX2 for instance, probably best to do development - on the development board and then transfer it to one of the carrier boards for actual implementation, so you still really need the development board in any case.

    Which really is a great price for what you get, this is the front running state of the art development platform - period.

    It does however require a serious investment of time and money.

    Best Regards,

    Gary

  • Hughes: Sure! Vision processing on companion computers, neural network implementations, swarming and tricky path planning, ... Lost of applications that require and can always use more processing power.

  • Well, personally I am interested in developing autonomous UAV capabilities, sense & avoid, collision detection, also realtime mapping with 4K cameras, ect.  Always need more compute!  The screenshot from OP is FCN-Alexnet segmenation network that I've been training on FPV imagery, now I'm adding landing targets, urban structures, industrial equipment, ect.  Jetson TX2 has enough horsepower to run it onboard in realtime.  I'm preparing to testing on Teal drone and Spiri developer drone.

  • MR60

    Do we really need a Xeon server performance within a UAV today or even in the next year?

  • Developer

    I get that the Jetson hardware is expensive for a hobbyist, but the way I see it they are not really made for that marked after all. I don't think aiming for your typical DIY price segment could justify the development cost of something like this. The performance power ratio on the TX2 is just insane.

  • The devkit carrier board is the open reference design that all the deployable designs are derived from, and needs to break out the features off the module, in particular the CSI camera header and multi-display expansion header, M.2 and SDcard slot, PCIe slot, SATA port, and others necessitate mini-ITX for the devkit.

    As a result of fully-featured reference carrier, Jetson has grown a plentiful 3rd-party ecosystem where developers have created a wide range of deployable options in short order. With so many options and use-cases in embedded, NVIDIA can't possibly begin to directly provide everyone's final breakout board. Even within drones, multiple varying options have sprung up - ConnectTech has 10 different embedded boards, and even a 1U array server with 24 modules. Rather than restrict the options of end users, we kept all the features of the Tegra SoC and designed Jetson module to provide 400 signals (this is why it's not an SBC). NVIDIA keeps focused on keeping the lead with our module roadmap and enables everyone else.

    Note that you can order bundles from ConnectTech (and probably Auvidea if you inquire) with just the module and their deployable carrier. And if any of you guys have an EDU email address, get the EDU discount.

    Regarding the software, NVIDIA's board support software is pre-integrated into frameworks like Caffe, Torch, TensorFlow, Theano, OpenCV, OpenVX, ect.  Rather than knowing the details of GPU programming, the open frameworks already leverage GPU acceleration out of the box. It's the best of both worlds. Jetson TX2 gives you the performance of Xeon server in a credit-card module...it blows Joule away, no questions asked. Joule is expensive for delivering more than 10X less perf and basically being more towards rPI domain - Jetson delivers the ultimate performance for embedded.

  • I'm with @Hugues, the tx1 has been too expensive and the dev kit is ridiculously large for diy-ers.  I've had my eye on it enviously ever since it came out but just can't justify the cost of it by the time you've added an auvidea carrier.  And most of the funky stuff is essentially proprietary.  I just bought an intel joule for half the price, and the dev kit is smaller than a raspberry - perfect for playing around with on a small drone.

    The TX2 looks awesome and will definitely be on my christmas/birthday list, but it needs to be more accessible.  If the modules are more easily available without the dev kit at a reasonable price, and cheaper carriers (i like the look of the sprocket!) then it will probably be more popular.

This reply was deleted.