Tesla is Building its Own AI chips for Self-Driving Cars

Discussion in 'HardForum Tech News' started by DooKey, Aug 2, 2018.

  1. DooKey

    DooKey [H]ard DCOTM x4

    Messages:
    7,822
    Joined:
    Apr 25, 2001
    Elon Musk announced that Tesla will begin building its own AI chips for self-driving cars during their earnings call yesterday. Until now they've been using Nvidia's Drive Platform, but they say their in-house chips will focus on their specific needs and will run faster than current hardware. Also, the new computers based on in-house chips will plug into existing cars because they have compatible connectors. Hopefully this works out for Tesla. You know this has to grate on Nvidia.

    By having its own silicon, Tesla can build for its own needs at its own pace. If they suddenly recognize something the hardware is lacking, they’re not waiting on someone else to build it. It’s by no means a trivial task - but if they can pull it off without breaking the bank (and Elon says it costs them “the same as the current hardware”), it could end up being a significant strength.
     
  2. darckhart

    darckhart Limp Gawd

    Messages:
    237
    Joined:
    Jun 15, 2013
    What Elon really meant to say:

    "So you know how we could never deliver on our production numbers? It's because we were spending all our resources building in-house AI chips."
     
    Armenius likes this.
  3. ChadD

    ChadD 2[H]4U

    Messages:
    3,693
    Joined:
    Feb 8, 2016
    This does make sense long term. Nvidias general compute platform/chips are not a long term solution for dedicated function tasks like self drive. Once you have AI trained... there is no need for programmable chips. A custom ASIC chip that does nothing but accelerate that trained AI... should be both faster, more power efficient, and more reliable.

    Self driving after all the AI training is done doesn't require 1000s of dormant programmable shader cores.

    With all the talent out in the world right now in terms of building ARM powered SOC... I don't think this is nearly as risky as some might assume. Licence ARM for the general CPU work which would be no different from then the Nvidia chips.... sandwich in some custom ASIC built to drive a specific trained AI Layout. Its the future of all things AI. Nvidias future is in the server, AI overlay creation space. Unless they enter into agreements to make more task dedicated chips. Self driving cars, drones ect... need more task oriented chips not general purpose compute cores, this should also drive the cost down allowing self driving tech to end up in less expensive drones ect.
     
    Wrecked Em and Spidey329 like this.
  4. Verge

    Verge [H]ardness Supreme

    Messages:
    6,113
    Joined:
    May 27, 2001
    *goes on facebook

    *makes 200 fanboi posts

    *lacks any understanding of what ARM SoC is




    Sure they "made" a custom processor...
     
    Armenius likes this.
  5. Johan Steyn

    Johan Steyn [H]Lite

    Messages:
    84
    Joined:
    Dec 1, 2016
    We all know this is what Keller was doing at Tesla. This has been awaited for very long. The most likely product to use, would be an AMD product, especially with Keller. It might be a special kind of APU, with many Vega cores. It could be a beast.
     
  6. pcgeekesq

    pcgeekesq [H]ard|Gawd

    Messages:
    1,403
    Joined:
    Apr 23, 2012
    The good news is:
    1) this puts pressure on NVidia to produce a better, more cost-effective self-drive platform
    2) is Tesla fails, they'll just buy the HW they need from someone - like NVidia.
    3) if they succeed, they may have an advantage over their competitors.

    I don't see how Tesla can lose much on this. NVidia isn't going to stop developing their self-drive platform, and they're not going to refuse to sell it to Tesla: possibly for less than if Tesla hadn't made this announcement.

    Like many public statements about future products, there's some negotiation tactics involved.
     
    Rahh likes this.
  7. Stoly

    Stoly [H]ardness Supreme

    Messages:
    6,108
    Joined:
    Jul 26, 2005
    Didn't they announce this like years ago?
     
  8. Twisted Kidney

    Twisted Kidney 2[H]4U

    Messages:
    3,503
    Joined:
    Mar 18, 2013
    I always wanted to know if bouncing babies were still bouncy as teens, sign me up to test your chip!
     
  9. Spidey329

    Spidey329 [H]ardForum Junkie

    Messages:
    8,677
    Joined:
    Dec 15, 2003
    Well, this little nugget may have worked. They announced record losses, yet their stock surged $49 (16.7%). Short sellers were squeezed by about a $1bil+ today.
     
  10. AlphaAtlas

    AlphaAtlas [H]ard|Gawd Staff Member

    Messages:
    1,713
    Joined:
    Mar 3, 2018
    Which is ridiculous, as Tesla's market market cap was already way more than Ford's IIRC (which makes no sense if you look at the volume both companies are putting out, and at the progress of their self driving and electric cars).

    That stock is way, way too inflated.
     
    Silentbob343 and Armenius like this.
  11. Gideon

    Gideon [H]ard|Gawd

    Messages:
    2,000
    Joined:
    Apr 13, 2006
    Hmm you ever get that feeling maybe Nvidia is not the easiest company to work with.
     
    katanaD and R_Type like this.
  12. HockeyJon

    HockeyJon [H]ard|Gawd

    Messages:
    1,057
    Joined:
    Dec 14, 2014
    That's because Tesla isn't a company; it's a religion.
     
    Silentbob343, Armenius and Unter Dog like this.
  13. Wierdo

    Wierdo [H]ard|Gawd

    Messages:
    1,776
    Joined:
    Jul 2, 2011
    Lotsa great news in the conference call, apparently they sold more Model 3s than all their competitors combined (defined in their chart below), and overall the 8th best selling vehicle in America despite the entry level luxury pricing.

    Also interestingly they said 4 of the top 5 non-Tesla trade ins were much lower priced vehicles such as Toyota Prius, Honda Civic/Accord and Nissan Leaf, indicating customers see enough value in the product that makes them jump into the entry level luxury segment.

    That sounds really promising for future demand, and this is before they even introduced the lower priced $35k baseline config.

    On the energy side they are constrained by trained electricians, demand is higher than supply, so most of their efforts are going toward large scale projects while training their consumer-side's workforce for the last two years.

    And it also sounds like they're slowly approaching level 3 autonomy with this coming month's software update, exciting times.

    DjjJ5bvUcAAveMG?format=jpg.jpg
     
    Last edited: Aug 3, 2018
  14. AlphaAtlas

    AlphaAtlas [H]ard|Gawd Staff Member

    Messages:
    1,713
    Joined:
    Mar 3, 2018
    To be fair, the chart is a little misleading. There's a ton of redundancy in the competitor's lineups that isn't shown.

    BMW, for example, has the 4 series as a coupe version of the 3 series, the 2, 1 and 5 series which are all somewhat similarly sized cars with wildly varying price tags, as well as some Z series models that are the same size. Off the top of my head, Mercedes has the CLA, CLS, E, SL, and SLK which are pretty similar... Same story with Audi and Lexus. Most have trims that hit that $50k price point.

    Tesla has the Model 3. The S/X are in a whole different price tier. Even if you eliminate the coupes, that chart is omitting a whole lot of the competition's competition.



    However, those still aren't bad numbers. And I think the 35K model is going to boost sales quite a bit.
     
    Last edited: Aug 3, 2018
    Wierdo likes this.
  15. M76

    M76 [H]ardForum Junkie

    Messages:
    9,002
    Joined:
    Jun 12, 2012
    Really, a little? Those 4 models are not the entire market. And let's see the numbers for vehicles actually delivered to the customers, not the number of people standing in line in front of their store.
     
    BB Gun and Armenius like this.
  16. mrp

    mrp n00b

    Messages:
    47
    Joined:
    Oct 28, 2005
    Custom processor from the guy that has concerns/fear of AI. Why am I thinking he will create Skynet one way or the other.
     
  17. Bowman15

    Bowman15 [H]ard|Gawd

    Messages:
    1,216
    Joined:
    Apr 7, 2015
    Bingo...might be a little pricey as well.
     
  18. R_Type

    R_Type Limp Gawd

    Messages:
    215
    Joined:
    Mar 11, 2018
    Microsoft will never go near them again either
     
  19. katanaD

    katanaD [H]ard|Gawd

    Messages:
    1,987
    Joined:
    Nov 15, 2016
    ok, so what they are really gearing up for is to be a new competitor in the console market...

    they announce they have games coming to their console, err i mean car.. and now their own chip

    so.. the new entry in the console market is high end self driving consoles, err i mean car

    NEAT !
     
  20. Armenius

    Armenius I Drive Myself to the [H]ospital

    Messages:
    16,826
    Joined:
    Jan 28, 2014
    Huh? Drive PX is dedicated to autonomous driving...
     
  21. Wierdo

    Wierdo [H]ard|Gawd

    Messages:
    1,776
    Joined:
    Jul 2, 2011

    Yeah one can count it in different ways for sure, but this is how Tesla chose to define it in this case:
    I tried to find a different breakdown of sales online, this is as close as I got to including stuff like the BMW 4 line in question, etc:
    4ca0b14673a898bea4891dd00c39009fbdd205fd772b8d7e5ebbb881303f840e.jpg
     
    Last edited: Aug 3, 2018
  22. Verge

    Verge [H]ardness Supreme

    Messages:
    6,113
    Joined:
    May 27, 2001
    That chart is awful, i've been seeing it everywhere. Technically TSLA sold those 16000 cars years ago. Putting them on a time based chart that doesn't include the entire lifespan is meaningless.
     
  23. ChadD

    ChadD 2[H]4U

    Messages:
    3,693
    Joined:
    Feb 8, 2016
    https://en.wikipedia.org/wiki/Drive_PX-series

    Of course they have built something for that market... but the chips are not custom built for ONE AI. They are still general compute cores.

    The issue is this. Nvidia is making hardware which is great for TRAINING AI. Once you have a AI neural network trained to do a task you don't need the same level of horsepower. At that point the absolute fastest way to do things is to build a ASIC chip that is built to push that trained AI and nothing else.

    Think of it in terms of your brain... you have billions of neurons, when your a kid learning how to do a task your brain lights up tons of areas a millions of neurons your don't really need for that task as your learning. However once you have learned how to do that task your brain remembers you only need this one area of your grey matter to do that. This is why biologist say this part of your brain is for X this part is for Y ect.

    AI works the same way while training the amount of horsepower required is high, the AI doesn't know what it will need yet. It figures out what it really needs to do the task. Once its trained it doesn't need all that horespower.

    Tesla, Waymo, and the major car companies can't go and buy a 15 thousand dollar Nvidia AI training kit for every vehicle they ship... that would be insane. It is also going to be less reliable... which is of course why Nvidia includes 2 chips in their drive hardware.

    The problem with Nvidias drive strategy long term... is they are building training hardware, and won't own any of the trained AIs that result from it. So imo they are going to have to strike secondary deals with the major car manufacturers and AI user to produce custom (or at least more specific built) chips to power their trained AIs. Either that or Nvidia is going to have to slash their prices by like 90% when it comes time for full production. The alternative is companies like Tesla who use their hardware initially and eventually build or contract their own custom chips elsewhere.
     
  24. ChadD

    ChadD 2[H]4U

    Messages:
    3,693
    Joined:
    Feb 8, 2016
    Well his custom chip won't be all that smart.

    I know I just used the analogy in my previous post... But think of AI chips as a human brain.

    Nvidia is offering billions of neuron complete brains, clean when you buy them. Ready to learn whatever you wanna teach it.

    When you train a brain to do something, the brain figures out which parts it needs to fire up to do that one specific task.

    Once the brain knows how to do that... it becomes route the brain stops firing all the extra neurons it was using when it was learning, cause it already knows the min amount of neurons it needs to fire up to accomplish the task.

    Tesla is talking about making a chip that doesn't even have the physical bits to learn anymore. It will simply have all the bits it requires to accomplish the task its been given.

    This makes the chip cheaper, faster... and importantly less expensive. If we are going to have fleets of self driving cards and trucks and drones. They need systems that cost 100s of dollars not 15k a pop.
     
  25. BB Gun

    BB Gun [H]ard|Gawd

    Messages:
    1,478
    Joined:
    Jan 14, 2004