NVIDIA Jetson Xavier NX Developer Kit

hello it's Jim from jetson XCOM today we're taking a look at the brand-new $399 Nvidia Jets and Xavier an ex developer kit in Didius and over HSN Xavier NX developer kit for us to play with let's see two pieces of tape one on each side let's open up the box that looks pretty promising let's pull the flap up we have the power supply this one's not for sale but you could buy one from the link in the description below 19 volts oh and it's a 45 water let's put that aside and we have the power cable or the power supply of course and then we have our board no wonder how you get out of here Oh yank here there we go and the board should slide right out arrey mmm here's a little card and the instructions QuickStart support guide anything else nope good to go gee let's put our boxes away flip it over attention observe precautions Oh we'd better ground ourselves okay we'll lift up the tape come out of there you little pig dog ooh that's interesting at first glance it's slightly different than a Jetson nano the carrier board is mounted on a plastic base plastic base all of the connectors look about the same 40 pin GPIO header PWM no I don't see a BOE header on here power over ethernet - CSI camera connectors same connector layout as the Jetson Nano the power jack is 19 volts there's a fan integrated into the heatsink I like that this only runs off 19 volts so we don't need a power jumper let's flip it over here's an m2 keee connector with a wireless card installed it looks like it's a standard wireless card from a laptop and here are two antennas for the Wi-Fi card that are captured by the plastic base and that's pretty cool and here's the m2 key Emma slot full-size this is nice to have you can rent something like a and vme SSD in it those are the immediately obvious things there are two screws that hold in the Jetson Xavier and next module and the familiar Kepler clips let's take out the module see what's underneath I removed the screws release the capture clips we can see the twelve pin button header here's the module itself it's pretty small the heatsink has fins on it it's a nicer thermal solution than the Jets nano and here's the PWM cable that goes into the carrier board to control the fan all in all I say a good first impression the Justin Xavier NX developer kit has two parts the first part is the Jetson module which holds the compute elements and memory and the other part is the carrier board which provides IO connectors for the system let's discuss the Jetson module first here's the picture of the Jetsons Xavier and X module the heatsink with the integrated fan has been removed this is generically called a system on module or som for short the Jetson Xavier NX module is available in two flavors the first is the development kit version which we are looking at today the other is what is called a production module the two versions are slightly different the dev kit version uses a micro SD card for secondary storage the production module has onboard emmc memory instead also the production module has a thermal plate instead of the integrated heatsink with fan people use the Jetson production model to create new products on the Justin NX module sit San Xavier Tegra system on a chip or SOC the idea behind a SOC is that by adding some external memory and some simple blue chips you end up with a complete working computer by integrating all of the compute functions into one chip you get much better performance and energy efficiency bottom line need compute just pop in a Jetson module the Jetson Xavier NX module has a 260 pin sodimm connector which mates with the Jetson carrier board the module measures 45 millimeters by sixty nine point six millimeters techSpec time here are some highlights of the Xavier NX we have a six core 64-bit Nvidia Carmel ARM version 8.2 CPU there are six megabytes of l2 and four megabytes of l3 cache the GPU is Nvidia Volta architecture there are 384 cuda cores and 48 tensor cores there are 8 gigabytes of 128 bits low-power ddr4 X memory and the memory throughput is 50 1.2 gigabytes per second the xavier NX runs in two power profiles 10 watt and 15 watt modes in 15 watt mode you get a eye performance of 21 tops using into 8 for speeding up inferencing there are two Nvidia deep learning accelerator engines there is a vision accelerator with a southern wavy liw vision processor the expected hardware video encoders and decoders are present along with a cortex and five sensor processing engine we'll talk about some of the low and high speed communication channels that the Justin NX provides when we talk about the carrier board flipping the board over we have the underside of our dev kit module you can see our microSD card reader and the memory chips bottom and top the cross-brace in the middle holds on the thermal solution let's take a tour of the carrier board this is a top side of the board the xavier NX module has been removed up here we have two mipi CSI to defy camera connectors these connectors will allow you to work with raspberry pi version 2 cameras right out of the box on the right side here we have a 12 pin button header this brings out system power reset UART console and force recovery related signals notice here that we have an unpopulated area of the board which appears to be ready for installation of the can bus header down here we have a 40 pin GPIO expansion header in a Raspberry Pi layout one thing that is different here is that there is no power over ethernet or Pio e connector like the Raspberry Pi has finally on the left side we have the main external connectors top to bottom nineteen volt power input hdmi and displayport connectors for USB 3.1 connectors Gigabit Ethernet and a USB 2.0 micro B connector let's look at these connectors from a different angle the display connectors are stacked DisplayPort on the top HDMI on the bottom there are two stacks of USB 3.1 connectors for all together here are the four screws that attach the carrier board to the plastic base the board measures 103 millimeters by ninety point five millimeters by 31 millimeters the development kit weighs about a hundred and seventy-six grams both of these are what the Jetson module attached let's flip the carrier board over here is a m2 key east lot populated with a wireless card this appears to be a standard laptop type of device the wireless card comes pre-installed on the kit notice that the antennas for the card are captured in the plastic base there is also a full-size m2 key M slot which is useful for adding extra storage like a nvme SSD let's talk about the Jetsons software we have our usual friends jetpack 4.4 si UD NN tensor RT deep stream and the Issac SDK these software libraries are the backbone which makes the Jetson a platform accelerator to speed up most if not all tasks in an AI pipeline here's the jetpack SDK page this is where you go to set up your Xavier NX for the first time NVIDIA provides many resources for learning a eye on the Jetsons hello AI world is a great introduction to a eye on the Jetson platform two days to a demo is a series of deep learning tutorials to deploy AI and computer vision to Jetsons in the field Jetson zoo contains instructions for installing various open source add-on packages and frameworks like tensorflow pi torch MX net and caris on Jetson and deep learning Ross nodes helps developers integrate AI capabilities developed with Nvidia's AI tutorials with robot operating system Ross the next big push on the Jetsons is cloud native containers powered by Nvidia jetpack SDK Jetson employs the same Nvidia CUDA X accelerated computing stack and Nvidia container runtime used in data centers and workstations around the world I just read you a marketing blurb haha those are funny the basic idea here is that NVIDIA has set up a server called NGC to distribute containers because Nvidia tools are all cross-platform applications can be created with the latest Nvidia tools and optimized directly on HSN developer kit a PC slash workstation with cross-compile toolchain or even in the cloud using cross compilation containers enough talk let's start this puppy up okay let's wire this baby up I want to share with you a couple of pro tips the first tip is to use a high quality as d card I use a Samsung Evo I've had good luck with those don't waste your time buying the minimum required size you want to make sure that you have enough room for development pros don't do the minimum required they go large the SD card reader as you recall is on the underside of the Jetson module it's here let's insert the card the pin side goes up press it in you should be able to feel it seat and hear a little click you're pressing against a little spring the next pro tip is to get out your label maker I have mine right here you want to label your power supply especially if you have more than one Jetson and if you're like me you can put on the labels slightly crooked so it can trigger your OCD every time you look at it the third tip is to put some electrical tape on the plug and the jack for the nanos I use green for this one I'm using white it makes it easier to tell which plug goes where if you use black electrical tape I think you've missed the point we will plug this in in a moment let's plug in our keyboard let's plug in our mouse and our HDMI cable will connect that to our monitor we are running the demo off of a Western Digital WD black high performance and the meas s de 250 GB s you know we love our GPS I will leave a link in the description below the moment of truth let's plug this puppy in we should see a green light go on there it is let's switch over to our screencast it's demo time the Xavier NX is all booted up this is the familiar Jetson in the home screen I've been using the Xavier NX for a little while now it feels very much like a laptop or a desktop in its performance we can mix and match how many cores we are using the supplies define performance and each one of the power envelopes we use Nvidia sent over a demo for us to explore it shows off the power of the Xavier NX everything in the demo is running from containers it takes a couple of minutes to set up be right back okay the demo has started up wow the fans just cranking there are southern models running simultaneously manage it's an N or RT x2 it just choke and die on this impressive in the top left corner we have the people detection container this is running a containerized people detection inferencing tasks using deep stream it's analyzing for concurrent 1080p video streams to identify the number of people in each stream this is a resonant 18 model input image size is 960 by 540 4 by 3 and the model was converted from tensorflow to tensor RT to run on the Jetson in the bottom-left corner we have pose estimation the container is running a neural network to estimate the pose of the people in the input video stream you can see where it tries to model where there are shoulders their hips legs arms and face are this neural network is a resonant 18 model with input image resolution of 224 by 224 the model was converted from PI torch to tensor RT in the bottom right corner we have a gaze estimation container you can see that when the subject looks directly at the camera the boxes around the person's eyes turn green you might use this as a cue to interact with the person I would probably use it to determine if they were shifty in the top right corner we have a natural language processing demo we call these in old P this is running the Bert NLP neural network it's pretty demanding is my understanding it basically takes questions through a voice input and provides relevant answers based on the passage that it's associated with this demo uses two models quartz net 15 X 5 for a speech recognition that is converted from PI torch to tensor RT and then we have the Bert based model for language model to NLP which was converted from tensor flow to tensor RT let's do a Bert demo that's trying to make sense of what it knows about Burt what is Burt it's a deep neural network 31.4% what does Burt mean by directional encoder representations from Transformers fifty nine point one percent what can Burt do it can perform a variety of NLP tasks hmm it seems pretty happy with itself how does Burt work it got earth instead of Burt this isn't as sophisticated as Siri or Alexa or Google but it's certainly the beginning and it all runs on device what is GTC it's the GPU technology conference let's see Jetson Xavier what is Jetson Xavier it is itself how many cores does a Judson Xavier have $399 how many CPU cores so it's a little bit off on that but more importantly let's talk about football when did the NFL season begin September 5th 94% who won the Super Bowl Kansas City Chiefs where did they play the Super Bowl hardrock stadium who are the Raiders ok the City Chiefs oh we better fix that there'll be riots in the street let's edit this where did the Raiders play in Las Vegas much better where were the Raiders Oakland in Los Angeles who are the San Francisco 49ers Kansas City Chiefs I knew it those people are all the same if you liked this video give it a thumbs up and if you have not already please subscribe thanks for watching oh we're supposed to say stay safe [Music]

Channel: JetsonHacks

Views: 75,126

Keywords: GTC 2020, Jetson Hacks, Jetson Xavier NX, NVIDIA Jetson Xavier NX Developer Kit, Jetson Xavier NX Developer Kit

Id: aUz1bNGm04w

Length: 21min 46sec (1306 seconds)

Published: Thu May 14 2020