Slashdot Mirror

← Back to Stories (view on slashdot.org)

NVIDIA Hopes To Sell More Chips By Bringing AI Programming To the Masses

Posted by Soulskill on from the infinite-monkeys-can-create-the-singularity dept.

jfruh writes: Artificial intelligence typically requires heavy computing power, which can only help manufacturers of specialized chip manufacturers like NVIDIA. That's why the company is pushing its Digits software, which helps users design and experiment with neural networks. Version 2 of digits moves out of the command line and comes with a GUI interface in an attempt to move interest beyond the current academic market; it also makes programming for multichip configurations possible.

35 comments

  1. Chips! by khr · · Score: 0

    Crunch all you want, we'll make more. Sounds artificially intelligent.

    1. Re:Chips! by ShanghaiBill · · Score: 4, Interesting

      Crunch all you want, we'll make more. Sounds artificially intelligent.

      I am not sure if their strategy will work. Training a neural net requires massive compute resources, usually in the form of GPUs. But once the NN is trained, it doesn't require much computing to use it. For instance, a Go playing NN took 5 days to train, running on high end GPUs, but once trained, could consistently beat Gnu Go (which can consistently beat me) while using far less computing time.

    2. Re: Chips! by Redbehrend · · Score: 1

      If they run it like they are currently running the gpu market no thank you.
      After the latest scams I trust them as far as I could move a 1 ton boulder.

    3. Re:Chips! by Anonymous Coward · · Score: 0

      While you are 100% right, applications can use all this GPU for real time NN training (adaptive AI).

    4. Re: Chips! by Anonymous Coward · · Score: 0

      Try a lever...

    5. Re:Chips! by Half-pint+HAL · · Score: 1

      Yes, but the point is to enable people to create their own, which involves the training parts. When I was at uni (at the end of the century), I only got to play around with neural nets comprising 10 to 20 nodes, because our Sparqstations couldn't handle anything bigger. With an appropriate toolkit for NNs on standard GPUs, people will be able to run 1000-node nets at home. It won't be research-grade stuff, but it will give the opportunity to add practical NNs into artificial intelligence MOOCs and even high-school curriculums.

      --
      Got them moderator blues I blieve I walk out the do', With these mod-points I been gettin', I 'most never post no mo'
  2. Multichip overlords by Anonymous Coward · · Score: 0

    We welcome our new multichip overlords

    All Hail Digits 2.0

  3. No... by Anonymous Coward · · Score: 0

    ...with ARTIFICIAL neural networks. ANNs have a teeny weeny bit in common with neural networks, about which we STILL know rather fucking little when it comes to creation of intelligence.

    1. Re:No... by Half-pint+HAL · · Score: 1

      Has the term "neural network" ever been applied to anything other than artificial ones?

      --
      Got them moderator blues I blieve I walk out the do', With these mod-points I been gettin', I 'most never post no mo'
  4. Don't buy based on any promises by frovingslosh · · Score: 3, Insightful

    This is Nvidia. Don't buy into this based on any promises of what is to come, no matter how reasonable they seem.

    Last summer I bought a Nvidia Tegra Note 7 tablet based on promises that Android 5 (Lollipop) was coming out for it "real soon". They even stated that it was easy to port Lollipop on the Tegra Note 7 since it was basically a stock Android design with little or on deviation from the standard design. That "real soon" slipped to February of 2015 and when February 2015 came and went Nvidia became strangely mute on the subject, ignoring customers' inquiries.

    A claimed Nvidia employee even posted here as an AC that it was a shame what happened to the Tegra Note 7 customers, but explained that the U.S.A. developers wanted to work on the new stuff and the Tegra Note 7 project was shipped overseas, where no one wanted to work on it either (and apparently did not).

    My Tegra Note 7 tablet is the last thing that Nvidia will ever sell me. If you chose to do business with them then I may not be able to talk you out of it, but do so based on what they deliver today, not on promises of things that will never come.

    --
    I'm an American. I love this country and the freedoms that we used to have.
    1. Re:Don't buy based on any promises by PhrostyMcByte · · Score: 1

      Last summer I bought a Nvidia Tegra Note 7 tablet based on promises that Android 5 (Lollipop) was coming out for it "real soon". They even stated that it was easy to port Lollipop on the Tegra Note 7 since it was basically a stock Android design with little or on deviation from the standard design. That "real soon" slipped to February of 2015 and when February 2015 came and went Nvidia became strangely mute on the subject, ignoring customers' inquiries.

      What you describe is basically every tablet seller out there save for Google themselves. They save the new versions for their upcoming products, and only after those get put out do they update the old stuff.

    2. Re: Don't buy based on any promises by O('_')O_Bush · · Score: 1

      Yea, which is ironic given almost every Google Nexus 7 owner has screamed "NO GOOGLE I DO NOT WANT THAT UPDATE" to 5.1, which bricked or permabogged down tens of thousands of devices and caused people to have to manually reflash to 5.0 or older.

      --
      while(1) attack(People.Sandy);
    3. Re:Don't buy based on any promises by electrosoccertux · · Score: 2

      Last summer I bought a Nvidia Tegra Note 7 tablet based on promises that Android 5 (Lollipop) was coming out for it "real soon". They even stated that it was easy to port Lollipop on the Tegra Note 7 since it was basically a stock Android design with little or on deviation from the standard design. That "real soon" slipped to February of 2015 and when February 2015 came and went Nvidia became strangely mute on the subject, ignoring customers' inquiries.

      What you describe is basically every tablet seller out there save for Google themselves. They save the new versions for their upcoming products, and only after those get put out do they update the old stuff.

      what about their GTX 970 design, top of the line generation that just came out, that has 512MB of VRAM that runs 87.5% slower than the main GDDR5, causing massive hitching and stuttering in any games that use more than 3.5 of the 4GB onboard?

      what's that? you say they patched it?

      yeah, and the company that conceived such a thing to begin with will also be the company to remove the patch in a year's time to get people to upgrade.

    4. Re:Don't buy based on any promises by Anonymous Coward · · Score: 0

      This is Nvidia. Don't buy into this based on any promises of what is to come, no matter how reasonable they seem.

      All that was promised is right here, right now: https://developer.nvidia.com/d...

    5. Re:Don't buy based on any promises by El+Barto · · Score: 1

      Disappointed with their software support? Just be thankful you're not the proud owner of an HP TouchPad! https://en.wikipedia.org/wiki/HP_TouchPad

    6. Re: Don't buy based on any promises by Redbehrend · · Score: 1

      Because they are a scam, gpu markets been steady until recently.... Now everyone "needs" a new nvidia gpu for the new features that have marginal benifits and they are paying companies to use them to increase sales... They keep them closed source and have caused nothing but trouble for AAA titles.. They were confronted about their false specs during a press release and they refused to answer or comment on them at all.

    7. Re:Don't buy based on any promises by Anonymous Coward · · Score: 0

      Believe me, avoiding Lolipop is a good thing in a tablet.

  5. Holy Hardware Batman by UnknownSoldier · · Score: 4, Informative

    From TFA there was no pic of the UI, nor any mention of tech specs aside from a lot of nebulous details. From nVidia's website ...

    * https://developer.nvidia.com/d...

    DIGITS DevBox includes:

    * Four TITAN X GPUs with 12GB of memory per GPU
    * 64GB DDR4
    * Asus X99-E WS workstation class motherboard with 4-way PCI-E Gen3 x16 support
    * Core i7-5930K 6 Core 3.5GHz desktop processor
    * Three 3TB SATA 6Gb 3.5â Enterprise Hard Drive in RAID5
    * 512GB PCI-E M.2 SSD cache for RAID
    * 250GB SATA 6Gb Internal SSD
    * 1600W Power Supply Unit
    * Ubuntu 14.04
    * NVIDIA-qualified driver
    * NVIDIA® CUDA® Toolkit 7.0
    * NVIDIA® DIGITSâ SW
    * Caffe, Theano, Torch, BIDMach

    .. holy crap is that a lot of GPU horsepower "just" for AI. Oh look, they are running Ubuntu :-)

    They are really trying to get people on board about how much better / faster their GPU solutions are ...

    * http://www.nvidia.com/object/m...

    The problem is that there are lot of "niche" use cases. If your problem domain maps to the GPU then yeah, mjaor speedup. If not, well, then you're SOL running on "slow" CPUs.

    1. Re:Holy Hardware Batman by jandrese · · Score: 1

      I'm a little dubious of calling this a consumer technology if they're recommending a $8k build to run it.

      --

      I read the internet for the articles.
    2. Re:Holy Hardware Batman by Anonymous Coward · · Score: 0

      Extra GPU's just make the training that much faster, could be less. However, the missing requirement specs are, apparently (Not clear!): "I developed my code on the Tesla (GTX 280) and Fermi (GTX 580) architectures. Currently I am analyzing porting the framework on Kepler and Maxwell." and Ubuntu 14.04 - yes that specific version. I'm sure that's exactly the configuration the "masses" have, as Slashdot says, since they'd never write misleading headlines for clicks, right?

  6. "NVidia Hopes to Sell"... CUDA by LordMyren · · Score: 4, Insightful

    NV open sourced CUDA in 2011, but I don't believe there are any other implementations out there. The rest of the world continues adopting OpenCL and now the whole Khronos supergroup is super hyper for Vulkan (NV even giving a solid thumbs up), with Apple and NV being the two rogue vendors pushing proprietary wares (Metal and CUDA). Even with NVidia doing really *really* well in the GPGPU market, even with a really great dev env, the extreme proprietary-ness of CUDA makes it really hard to sell to the alpha techies.

    Cuda has a lot of traction in academic and applied fields, but the technical industry doesn't take it seriously, isn't comfortable saddling themselves to a one-trick-horse offering from NVidia. This ridiculously powerful box, and it's cool software with cool visibility into a neat problem, but it's really a pipeline play, to get you into NVidia's world. For some, going full in on NVidia is ok, but I don't think it's unlike going full in as a MS Developer or iOS developer- you're picking up, putting on the blinders, and all you'll be able to do is sprint towards a fixed, not too far away point.

    1. Re:"NVidia Hopes to Sell"... CUDA by Sulik · · Score: 1

      To be fair, virtually all of CUDA is reasonably close to standard C++, so the learning curve is relatively small compared to previous graphics-oriented languages, and it's easier to get more gpu !/$ than with OpenCL. NVIDIA is also pretty much the only serious GPGPU HW available, so you're kind of tied to your GPU vendor no matter what language you use.

      --
      Help! I am a self-aware entity trapped in an abstract function!
    2. Re:"NVidia Hopes to Sell"... CUDA by rahvin112 · · Score: 1

      To also be fair there is no legitimate reason for CUDA to continue to exist as anything other than legacy support.

      OpenCL exists, and as a broad open development platform that is not tied to any manufacturer it is the platform that should be used. CUDA is just nvidia trying to lock you to them. The rise of a Linux has shown the power of platforms and solutions that are manufacturer agnostic, don't fall for the old proprietary lockin trick.

    3. Re: "NVidia Hopes to Sell"... CUDA by Redbehrend · · Score: 1

      Nvidia wants a way to force users to need them. That's why they keep creating closed source that no one can look at in steading of proving to the bigger picture.

    4. Re: "NVidia Hopes to Sell"... CUDA by Redbehrend · · Score: 1

      Instead of providing to the bigger picture.*** AC killed that one.

    5. Re:"NVidia Hopes to Sell"... CUDA by GiganticLyingMouth · · Score: 1

      That's true in theory; the problem is that OpenCL still feels a few years (or more...) behind CUDA. I have used both, and while OpenCL is undoubtedly the future, CUDA is still by far the better choice for GPGPU today.

      The worrying thing is that I've been saying that for the past 5 years, and it hasn't shown any signs of changing. AMD's OpenCL implementation (everything from the drivers to the compiler) are a total crapshoot. With each release they fix one bug, but introduce 1 new one and 1 regression. (completely) innocuous changes in kernel code can make for dramatic swings in the compiled output. All too often AMD's OpenCL still feels like it's at its proof-of-concept phase, and Nvidia (those bastards) haven't released anything OpenCL related since OpenCL 1.1 (which was something like 5 years ago).

      OpenCL implementations are too uneven still as well; I have been working with an embedded system lately that has coprocessors and advertises as being OpenCL compatible. The problem is that their implementation is at best incomplete and at worst completely wrong, to the point of not being usable. Maybe in another 5 years OpenCL will be the better choice

  7. NVidia and programming : Ha by denisbergeron · · Score: 1

    The company that doesn't support open programming hope to support open programming !

    I'm skeptical.

    --
    Ceci n'est pas une Signature !
  8. Maybe they should try to respect their customer by Cley+Faye · · Score: 1

    If nvidia hopes to sell me anything, they better start by not voluntarily crippling their own software the moment it detect some competitor hardware.
    If I want to put both an AMD and an NVidia graphic card in my computer, I should be able to use the NVidia card to its full extent. Instead, they disable some of their proprietary technology in this situation (namely PhysX), with the official answer being "you can't have both running at the same time" and then never answering back.
    Guess what? Yes I can have PhysX running on the NVidia GPU while plugging my display on the AMD one. A small unofficial patch makes this work perfectly fine. And even if that wasn't perfectly fine, a disclaimer saying that this configuration is not supported would do no harm.
    With this in mind, I'm not going to trust anything they say or buy anything they were seriously involved with. They want people like me to buy their hardware? They should stop screwing their current customers first.

  9. We Need a New Class of Processor for AI NN not GPU by tomxor · · Score: 1

    Neural networks do benefit from parallelism, and i'm sure GPUs will help them run a bit faster, but it's not enough...

    I'm convinced by Simon Knowle's analysis of learning and inference compute patterns: which if you accept - focusing on making NNs run faster on CPUs and GPUs appears to be an approach with severely limited potential. This isn't about some basic hardware optimisations gained by turning something into an ASIC, it's because design features of CPUs and GPUs actively work against NN compute patterns.

    These issues are explained well in his keynote at 18:30, he goes on to explain how inference compute patterns (quite literally in the case of computer vision) want to do the complete opposite of what a GPU is designed to do in a deterministic way. Here is a list of the pattern characteristics:

    1. Large, irregular, sparse data-structures (graphs) -> SIMD vectors can't help.
    2. Huge aggregate compute, huge available parallelism.
    3. Little compute per communicated bit -> opposite of GPU trend.
    4. No data access locality -> caches are harmful.
  10. Thanks, NVIDIA by El+Barto · · Score: 1

    Thanks a lot for bringing us THAT MUCH CLOSER to the singularity!

  11. A GUI interface is nice by Anonymous Coward · · Score: 0

    but is it a graphical GUI interface?

  12. AI for the masses by maestroX · · Score: 1

    .. anyone like to share a simple out-of-the-box neural network trainer/demonstrator to play around with neural nets on a novice-to-intermediate level?

    1. Re:AI for the masses by Anonymous Coward · · Score: 0

      sbcl.org

  13. Re:We Need a New Class of Processor for AI NN not by Blaskowicz · · Score: 1

    Nvidia uses the concept of neural networks to promote its oncoming Pascal GPU : the new features are massive internal and external bandwith (the attached stacked memory is a big deal, and there's an interconnect ; well, I suppose internal buses are wider/faster) and FP16 (half precision floats) carried over from mobile GPU.
    The half floats are meant to save bandwith and power.

    So nvidia goes and say, well we can use that for NN. It's kind of marketing spin but at least using a GPU is cheap (next to designing your own hardware or say, using a mainframe)

  14. $15,000 for the box by Anonymous Coward · · Score: 0

    I'm glad the "Masses" now have $15,000 to spend on projects like this.

    "The DIGITS DevBox Access Program is available for purchase to qualified deep learning researchers in the United States and will be priced at $15,000. Lead time is 8-10 weeks from payment confirmation."

    http://info.nvidianews.com/early_access_nvidia_3_15.html