NVIDIA Jetson Xavier NX development kit review

This new development kit allows you to to use the NVIDIA Jetson Xavier NX System-on-a-Module (SoM) as a Single Board Computer (SBC). The Xavier NX comes with a hexa-core NVIDIA 64-bit CPU, a 384 core GPU, 8GB of RAM, HDMI, USB 3.1, and access to an SSD via an M.2 NVMe socket.
Article: www.androidaut...
Introduction to Android app development: www.dgitacadem...
Let Me Explain T-shirt: teespring.com/...
Twitter: / garyexplains
Instagram: / garyexplains
#garyexplains

[@santoshvasaresearch419]

There is a ton difference between TensorFlow and tensor cores. There's nothing called TensorFlow cores on this board. TensorFlow is a software deep learning framework used to build neural networks easily and is developed by Google. Tensor is nothing but a multi-dimensional matrix and tensor cores are cores that are specially developed by Nvidia to accelerate tensor computations(all these cores have to do is single multiply-add operation facility per core). I am a fan of android authority's content as well as your explain series and I like how you made everything simple here for people! thanks and love from India.

[@DMonZ1988]

you're so knowledgeable! i'd really really love to see more detailed tutorials, courses, project guides and real world applications of these things from you; i'm sure you're doing some incredibly interesting things and this stuff is fascinating! thanks for the cool video!

[@1MarkKeller]

*GARY!!!* *Good morning Professor!* *Good morning fellow classmates!* Stay safe out there everyone!

[@GaryExplains]

MARK!!!

[@andersonfrans]

15W Mini Desktop Gaming PC, what a 2020 year.

[@billykotsos4642]

That's my jam. Let's go. ML and Hardware

[@bertbrecht7540]

Thank you so much Gary! You just taught me a lot and also saved me $400, at least for now. 4x price over the Nano for average

[@Phynix72]

New Nvidia sheild tablet is a must...

[@MagDag_]

Good board, but it's old architecture VOLTA, a new one is AMPER which is much faster and less power-consuming. Latest boards working on the r platform.

[@ibonitog]

It's Titan V not 5. The V represents the Volta architecture not, that it is the fifth generation Titan (or something). In fact it isn't. No worries, just to let you know. Great video! :)

[@noridzwannordin6416]

"Titan ten"

[@esra_erimez]

It is not for Teen Titans?

[@ronaldronald8819]

Hi Gary Thanks for your review.

[@GaryExplains]

Very welcome

[@soraaoixxthebluesky]

4k resolution on 15W system. Wow

[@ngroy8636]

It's so much work to get things working on it. Arm is simply now well supported yet.

[@GaryExplains]

I don't agree. Please give an example.

[@ngroy8636]

Although Nvidia provides support for modules and packages like TensorFlow, Nvidia-docker, and PyTorch; and those are not officially supported can be replaced by community alternatives (such as miniconda, jetson-xavier-tensorflow-serving). Regardless, I stand by my statement. It is troublesome to get things working on jetson. For example, to install a python package such as Numba, I first have to compile LLVM and figure out how to install packages that are not available in the package manager before installing the package. Compared to installing Numba on x86-64 system, where everything is well supported and can be done with just pip install Numba, installing packages on jetson(and other arm64 systems?) can be time-consuming due to the dependency issues.

[@fuzzybabyducks7878]

NX only came out in mid april. it worked on the Nano devkit board. It was not available before then and only available for about 3 weeks before the devkit was released.

[@nerrufam7105]

The mos important question: How does emulation work on it? 😛

[@jimmymac2292]

For $400 it better emulate just about everything

[@JERZEELOON]

Right. If it ran Android TV or someone ported shield tv's os it could probably be an awesome emulation box

[@BayBye2011]

@@JERZEELOON but why buy this just get the Nvidia Shield TV Pro....and save a bunch of money

[@JERZEELOON]

@@BayBye2011 I have one but I'd buy this as well it's quite a bit more powerful

[@DenGuleBalje]

Probably fairly well, but not as well as a Snapdragon 855/865/+. The GPU on this thing is amazing, but emulation for the most part needs CPU power

[@jezzaboi2168]

I know that the jetson wasnt made for being put in a tablet, but I still think a linux tablet powered by one of them would be great

[@theofficialgman]

It’s called the nintendo switch. Can run linux pretty easily with the original ones. I run ubuntu on mine.

[@benhager2559]

I'm mostly interested in this machine to play games. I think there is a huge market for it, If they offer an operating system that has good working drivers for sound and video. I'll buy one if I can game on it

[@glenswada]

Extremely good rundown. Didn't download your prime test but looking at rust prime benchmark between nano and xavier I am guessing the $400 NX is about 25% faster than $70 odroid N2 (my current desktop machine). So unless AI or GPU are primary requirements, it ends up a mighty expensive desktop in comparison to cost/performance of other systems. Think I will wait for release of RK3588 before considering my next upgrade. Thank You

[@JayzBeerz]

Great info thank you Sir.

[@GaryExplains]

Glad it was helpful! But did you have time to actually watch all of it before you wrote your comment! 😉

[@tusharkathuria9]

@@GaryExplains Video says uploaded 9 min ago... Comment says 9 minutes ago..

[@JayzBeerz]

@@GaryExplains yes lol. I think i'm gonna have to buy one now.

[@GaryExplains]

@@tusharkathuria9 The comment was posted within a few minutes of the video going live. 🤷‍♂️

[@1MarkKeller]

@@GaryExplains Everyone knows to comment BEFORE watching the video Professor! The classroom is getting bigger so you gotta act fast before our comments get buried in the avalanche.

[@williamhart4896]

I'd love to see nvidia drop out a 10 inch two in one shield tablet with that soc

[@mrmagootoo]

9:39 12.5 million primes said, 1.25 million shown. Typo or speech error? :)

[@GaryExplains]

Does it matter? The point is that the performance figures are relative as they all did the same task.

[@ferrarione4708]

Can you please do a video of using this board as a gaming pc

[@MaiONerds]

20 years later: There was even a dedicated CPU Called GPU to help run the graphics!!!

[@mattizzle81]

I found the Jetson Nano and Raspberry Pi 4 to be too slow and limited for my liking, maybe this would be a bit better. Worth a try.

[@TunjungUtomo]

well it's around 4x more expensive than the Jetson Nano, it's SUPPOSED to be better

[@pedrorotoli649]

Titan 5? Don't you mean Titan V, for Volta?

[@wrgknows]

Can you boot from the NVME?

[@zabique]

Question is can this run Deep face lab?

[@jeffg4686]

Could it be a VR headset controller? Like attached to belt or whatever with sata cable to headset. With AI at the very very edge (on person)

[@kevinpitts3548]

How does this favor the nvidia shield? Could you install their android TV software on it?

[@Katzelle3]

Nintendo Switch 2 right there.

[@TheRedFoxMcCloud]

no that's not gonna be enough but this would work for a switch pro

[@capricornone2362]

I predict something from Nintendo in 2021🤞

[@Katzelle3]

@@TheRedFoxMcCloud Spec wise it matches an XBox One S.

[@grima0482]

Does the Nvidia Carmel CPU also support armhf/ARM 32 Bit? I would like to install a 32 Bit armhf chroot with box86 on the Jetson Xavier in case I buy it to emulate some x86 games.

[@bschlueter]

Ssd Boot support?

[@marouanechouhou8312]

What is the difference between the GPU and the DLA? When using one or an other?

[@malathomas6141]

This versus udoo bolt v8?

[@GaryExplains]

Yes, that would indeed be a good comparison to make.

[@scudsturm1]

a titan 5, five, u mean Titan V(olta)

[@elgs1980]

My understanding is those dancing people examples do not make much sense on performance, because they are using trained models. I wonder how fast the training time is compared to a desktop gpu.

[@GaryExplains]

Of course it is using trained models, all ML applications use trained models. Also don't forget that decoding the video in real time and using the model is the key here, not just the raw inference speed.

[@elgs1980]

@@GaryExplains I think decoding video in real time is still much cheaper than the training. I wonder how this board compares to desktop gpu on training speed.

[@GaryExplains]

Training is always the hard bit. This board is designed for end products, for deployment. Training is of course possible using it, but that isn't it purpose. Secondly, of course, inference and even real time video encoding are "cheaper" than training, but the difference is that training happens infrequently and offline, whereas inference happens frequently and online (in real time). You care comparing apples with oranges.

[@GaryExplains]

Also, if you look at my Jetbot video you will see how I used the Jetson Nano for training.

[@ragilmalik]

what the hell is arm v8.2 ???? how does the architecture differ than armv8 ?

[@slappymcphee]

The Nano CPU limited 1.5Ghz was a BIG disappointment especially when it is a 2Ghz CPU. The JN actually pulled a lot of unsuspecting newcomers to the tinker space and them ending up putting them into a drawer to collect dust or selling them. Sad really that it happened the way that it did. The NX once again looks like a great entry. Just too bad that there is the $400 price point.

[@mattizzle81]

LMAO. Exactly. 😂 I got overly excited and bought 2!!! Even got 3D printed cases for them. Then when i started playing around with them the disappointment set in, and they are just sitting in a drawer waiting for a worthy replacement 🤣. To be fair my raspberry pi 4 is doing the same. So underpowered as to be useless for me. A smartphone offers so much more.

[@yourtubernoob]

@@mattizzle81 willing to sell one of those disappointments?? lmao I wanna check out one butt just blew my cash on a fully kitted Autel EVO 2 Pro XD

[@ngroy8636]

Takes forever to compile on Nano. P.s. I noticed there are two disabled cpu cores(?)

[@ruinoushavoc256]

I am looking for an encoding solution for OBS broadcasting. Even if I have to modify code, build from source or make my own solution this looks like my best low cost secondary Twitch streaming PC that can be scaled slowly into a cluster. Am I barking up the wrong tree?

[@Panhama]

nintendo switch 2 confirmed

[@power-max]

Can it run crysis? Or at least some sort of gaming?

[@sinchman1]

Is this device as powerful as a PS4 or XBOX1

[@ilyasahmad1191]

how to boot jatson nano to laptop

[@raventan1227]

Can you provide a link to replicate the 4x container demo on my NX? Massive thanks!

[@raventan1227]

Also, if you can recommend the dev board for NX that would be awesome!

[@dante19890]

this needs to be in the next nintendo switch

[@a.thales7641]

The successor needs yeah, not this. Cortex A78 + 512 Cores on 5nm or something like that.

[@aravindhanravichandran1089]

Can we use it for AR, VR application?

[@ClamGuy23]

This is going to be in the Switch Pro, make a video about it and get your views. You're welcome :P

[@MaiONerds]

Bro are you ok you kinda look pale/yellowish??? Man stay safe 🙏

[@GaryExplains]

LOL, I am using a new camera and light setup and I haven't get it 100% right yet.

[@MaiONerds]

@@GaryExplains ♥️♥️♥️

[@undisclosedmusic4969]

What is a TensorFlow core???

[@NERO-ez1mn]

can you make a sample used on it from programming for noobs who wanted to learn how that sh*t thing works

[@favesongslist]

This may be worthwhile checking out kzbin.info/www/bejne/gqa6iouBp7aWgJo

[@esra_erimez]

Is Xavier NX compatible with the Turing Pi board?

[@GaryExplains]

No.

[@PrivateSi]

I am pleased we a are seeing more compute methodologies being catered for directly but it's going to get very problematic..... We need a small, effecient but still extensible, standard low level and basic intermediate level instruction set combined to tie together the growing list of architecture extensions and addons being developed by many different companies... OpenGL/CL/AL DirectX (+ AI extension), Vulcan etc. are too high level and directX is not an open standard.. Risc 5 is an option but it has chosen a bloated, inefficient instruction format, unfortunately. The situation is going to get worse and worse. Defacto standards due to monopolies and market dominance are a result of lack of competition which is not good for consumers.. Still, it's good we're breaking free from traditional programmed expert systems and leveraging ever cheaper, smaller bulk computing to provide AI features.

[@DewsySipos]

How about training?

[@fuzzybabyducks7878]

Its okay... Ram speed is pretty slow in comparison to discrete graphics cards which has a couple issues, but i had two working together over the network with NCCL on pytorch and it was able to train faster than a single rtx2070 at the same price of 2 of these boards, but at 4x less power consumption. Its great for inference though. We have an array of 12 of them running in a 1U rack doing some inferencing for our company and they perform the job _much_ better than machines that cost 3 times the price and 5 times less power consumption. Edit: I should note, these were production NX chips running on jetson nano devkit boards.

[@umka7536]

Mmm. Is not it [Havier] like in Spanish?

[@GaryExplains]

What did they call Professor Charles Francis Xavier (AKA Professor X)?

[@internetjunkie5241]

Is this a raspberry pi killer when it comes to emulation? #AskingForAFriend

[@saltima]

RPi 4+ with GPU and NVME please

[@haijiechen2068]

价格让我死心