Slashdot Mirror

← Back to Stories (view on slashdot.org)

NVIDIA's $99 Jetson Nano is an AI Computer for DIY Enthusiasts (engadget.com)

Posted by msmash on from the how-about-that dept.

Sophisticated AI generally isn't an option for homebrew devices when the mini computers can rarely handle much more than the basics. NVIDIA thinks it can do better -- it's unveiling an entry-level AI computer, the Jetson Nano, that's aimed at "developers, makers and enthusiasts." From a report: NVIDIA claims that the Nano's 128-core Maxwell-based GPU and quad-core ARM A57 processor can deliver 472 gigaflops of processing power for neural networks, high-res sensors and other robotics features while still consuming a miserly 5W. On the surface, at least, it could hit the sweet spot if you're looking to build your own robot or smart speaker. The kit can run Linux out of the box, and supports a raft of AI frameworks (including, of course, NVIDIA's own). It comes equipped with 4GB of RAM, gigabit Ethernet and the I/O you'd need for cameras and other attachments.

13 of 127 comments (clear)

  1. Finally a board with some RAM by drinkypoo · · Score: 4, Interesting

    4GB puts this into the category where it's actually useful for stuff like web browsing. Sadly, the link to the item from TFA is 404, but it looks like it's actually got enough ports on it to be useful for doing stuff without needing a hub, too. Forget building robots with it, you can build kiosks. Do they have an Android build for it?

    --
    "You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
    1. Re:Finally a board with some RAM by Shaitan · · Score: 3, Interesting

      Sad but true and there really is no good explanation for it. I just closed everything an opened chrome, clocked in at 350mb. Loaded slashdot as the only page, suddenly it skyrockets to 850mb and then slowly settles back to 550-650mb. Now there definitely isn't enough content in this page to explain 10mb and even with the linked pages you have nowhere near the 200mb it has absorbed in content. It does make me wonder just what the hell it is using so much memory for.

      Compared to the 4mb footprint of Netscape 4 or IE 3.5 you've gained what... some adjustments to javascript and css? That explains maybe 10mb of the increase. A gargantuan cache? That would explain another 64mb maybe. The actual page content? If anything the pages actually have less content with html, css, and light weight icons being the styling of choice these days it's almost all text. A 2mb page would be massive but they use shitty autochurned output for most sites these days so call it 8mb, across all the linked pages that will fill your 64mb cache. With a little breathing room added in that is what about 100mb that can be explained?

    2. Re:Finally a board with some RAM by Shaitan · · Score: 3, Interesting

      It takes A LOT of pictures to explain 2GB. Good luck coming up with something in browsers that explains their memory consumption. Even with the... dear god 3600 lines of text that composes this page (vs the 50-75 you could do it in, maybe 150-200 if you were counting js and css). There just isn't any magic here. Even that ugly video enabled ad that is probably larger than the entire browser should be would only explain a few mb. This page and every linked page still isn't going to explain more than 100mb and the entire browser should itself should use less than that.

      I blame people obsessed with OOP

    3. Re:Finally a board with some RAM by quenda · · Score: 3, Funny

      Try searching for "4gb kodi box". Then narrow down to those with an SoC that has decent drivers for your OS of choice.
      Are they any good for you?

  2. mini-computers by Anonymous Coward · · Score: 2, Funny

    mini computers can rarely handle much more than the basics

    Minicomputer.

  3. I'll wait for the positronic version by nospam007 · · Score: 4, Funny

    This one doesn't even have the 3 laws.

  4. Mr. Fusion by theCat · · Score: 4, Interesting

    Soon people will be tinkering with personal-sized AI like they started to do with Arduino a few years back, and 3D printing more recently. The trend here is obvious, but we cannot predict what tinkers will come up with once they get their hands on these things in a big way.

    AI researchers fret about the "containment problem", meaning how do you prevent an autonomous intelligence from breaking out of your lab and doing whatever it wants to, including enhancing itself exponentially. So there is talk about creating process and protocols to contain AI similar to what you might have regarding biological containment for a microbiology lab working with dangerous pathogens. But those rules aren't going to work when anyone wants to can build a reasonably powerful AI machine using off-the-shelf components, and/or using cloud-based resources.

    I don't expect this is going to work out the way we think it is.

    --
    =^..^= all your rodent are belong to us
    1. Re:Mr. Fusion by drinkypoo · · Score: 2

      But those rules aren't going to work when anyone wants to can build a reasonably powerful AI machine using off-the-shelf components, and/or using cloud-based resources.

      Reasonably powerful? Do you mean capable of reason? Because it's likely that such will happen in a supercomputer before it happens in some hobbyist's garage.

      --
      "You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
    2. Re:Mr. Fusion by Kjella · · Score: 4, Insightful

      Reasonably powerful? Do you mean capable of reason? Because it's likely that such will happen in a supercomputer before it happens in some hobbyist's garage.

      Well maybe, maybe not. I mean we didn't start out as humans, before that we were monkeys etc. all the way back to single celled organisms. It might be that we're trying too hard to replicate intelligent behavior rather than find the underlying principles of intelligence. For example take AlphaZero, it was based on AlphaGo Zero which played Go. But rather than being a dedicated engine with a lot of specific programming for Go it just as easily beat the best at chess and shogi.

      Uber(!?) managed to pull off some impressive results with Go-Explore, a family of so-called quality diversity AI models. Open AI Five is constantly improving in DOTA, even though they're not playing the full game yet it's an open map with team play. Tencent is showing off some good Starcraft II play, even if they can't beat the best. DeepMind made some ass kicking Quake III Arena capture the flag bots. But is there some kind of supreme overlay that could evolve into all these AIs? I would say probably yes.

      That's really the holy grail of AI, what is like the "spark" of intelligence that learns to learn. It's probably relatively simple once we can look at it in retrospect, it's not a lot of code to do one specific task. It's some kind of general pattern to create more complex, dedicated "sub-AIs" for specific tasks like we have different brain centers in the brain. That and a few billion years of evolution, but computers can get pull that off pretty quick if we can just figure out where we're supposed to start. It'll probably start by beating us at tic-tac-toe, not any of the above.

      --
      Live today, because you never know what tomorrow brings
    3. Re:Mr. Fusion by del_diablo · · Score: 2

      I don't really agree.
      At the moment all these 'AI' are not AI. They are computer code running a comparison between a set of data and whatever is the realworld application of that data.
      So in the field, 'large advances' are when you figure out how to feed it a new type of data, and use it for a new type of task. Now Google has been doing this to satelite data for a few decades, to create realworld map data for consumption. But a newer development is that somebody figured out how to feed a comparison program with different resolutions of images, and then cobbled together image editing tools and upscaling algorythms so it could make A become B, and then apply that process.
      Which means the process involves a lot of bruteforce computing and comparisons.
      This recent example is waifu2x, which currently has a github page. And its been forked a few times, since waifu2x has a database and upscaling set intended for manga pages in different sizes. Now, it has been forked since then, into things like ERSGAN or AI Gigapixel, which do the same thing, but with slight alterations to the base image training set, and alterations to the tools uses to achieve the process

      But this is still basically a CRC machine, except the sensors can tell what size its working on.
      Interesting? But not really a AI per say, because so far its not even been capable of processing information into new information. Its basically looking at apples and oranges, and then using the datset of difference to give meaningful data if you input a durian but not b banana.

  5. Progress by JBMcB · · Score: 4, Funny

    30 Years Ago:

    Motorola's new 16-bit microcontroller has a whopping 16K of RAM, and 8 GPIO! You can use it to build robots, or home automation systems, or low-end general purpose computers...

    Today:

    Our new quad-core CPU, 128-core GPU is great if you want to build a speaker!

    --
    My Other Computer Is A Data General Nova III.
  6. Beowulf cluster by edi_guy · · Score: 2

    How can I make a Beowulf cluster out of these?

  7. This is just awesome by WaffleMonster · · Score: 2

    This is cool shit for what it is.

    What else has the ability to encode 4k h.265 in realtime /w comparable GPU at a price anywhere near what this thing costs?

    Having said that I do a lot of h.265 encoding and wouldn't touch the NVidia GPU encoders with a 39 and a half foot pole. They suck ass.

    Still at $100 the deal breaker will be what kernel and hardware support look like for this thing.

    Personally also looking forward to the N2.
    https://www.hardkernel.com/blo...