TMC is an independent, primarily volunteer organization that relies on ad revenue to cover its operating costs. Please consider whitelisting TMC on your ad blocker and becoming a Supporting Member. For more info: Support TMC

Tesla Autonomous Driving H/W

Discussion in 'Tesla, Inc.' started by scaesare, Aug 28, 2018.

  1. scaesare

    scaesare Well-Known Member

    Joined:
    Mar 14, 2013
    Messages:
    7,324
    Location:
    NoVA
    Opted to post my reply to @Krispykreme here for dear of retribution in the Investor thread...

    It may not make as much sense out of context, but here goes:


    Despite your condescending salutation to @mongo and claims of being in the autonomous driving industry, it appears your grasp on some of these things may not be up to snuff.

    Next gen HW is almost always consumes less power per flop than previous. Combine this with using specific-purpose compute cores (likely tensor processors), rather than more general purpose GPU cores, and your order of magnitude assertion is likely off target by a significant factor.


    Memory may not even need to be substantially larger. The memory size needs to be large wnough to hold the bueral-network. Just because the module has drastically more compute horsepower doesn't necessarily mean the memory size has to increase correspondingly. Heck, it could be the SAME size as the current design.


    Perhaps your seeming bad take on this whole situation is assuming this is a scaled up GPU platform. It's almost assuredly a tensor CPU (TPU), which is what Google and the other big-boys have done. A GPU is better than a CPU at a number of NN chores, but a TPU is yet in a different league.


    Custom silicon (an ASIC) can be a variety of designs... and even incorporate multiple architectures on the same die. Memory like SRAM can even be incorporated.

    There may very well be licensed CPU core as part of the design, but it's not what will be doing the heavy lifting. The real horsepower (and cost) is going to be the NN processing, again likely TPU design. And if you hadn't been aware, Jim Keller who designed this beast for Tesla is perfectly capable of a "from scratch" design...

    It's likely there are NO GPU licensing fees at all.


    But relatively cheap as compared to the NN processing portion of the system.



    Good... that matters for probably about 5% of the overall system.


    I've yet to see you make a compelling case for why it's needed. Dad.



    You seem to not have a real grip on things yourself. It kind of reminds me of the old guard who insisted that you needed specific hardware for everything all the while folks went ahead and didn't it in software anyway and they were left standing in the dust wondering what happened.

     
    • Helpful x 2
    • Like x 1
  2. lunitiks

    lunitiks Cool James & Black Teacher

    Joined:
    Nov 19, 2016
    Messages:
    2,708
    Location:
    Prawn Island, VC
    Gotta love invisible ink
     
    • Funny x 2
    • Like x 1
  3. scaesare

    scaesare Well-Known Member

    Joined:
    Mar 14, 2013
    Messages:
    7,324
    Location:
    NoVA
    #3 scaesare, Aug 29, 2018
    Last edited: Aug 29, 2018
    So apparently, my cut & paste mixed with the dark color scheme I used resulted in unreadable text if you are using the default light color sceheme.

    Whoops :p

    So...

    1) You can highlight it all with your cursor to see it

    B) Change your color scheme to dark temporarily

    III) A mod could help a brutha out now that my edit time has expired

    fore) Ignore it all together, as I ususally babble incoherently anyway....


    Honestly, the last option is probably best.
     
    • Funny x 2
    • Helpful x 1
  4. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,998
    Location:
    Michigan
    Autonomous driving doesn't really reasoning to work/ be safe. It needs a physics model along with a good object detection and movement detection system. To drive well, it needs to predict other vehicle types and movements (trolleys for instance).

    All radar use return intensity along with timing. Pulse return timing gives distance from the emitter/ receiver pair. Radar return strength is dependent on angle of target and physical/ electrical characteristics. Rough surfaces give a more omni-directional return vs something like calm water which give almost none unless you are perpendicular. Ability to penetrate objects and range resolution are dependent on center frequency and bandwidth respectively. A higher frequency and wider bandwidth pulse would resolve both the large and small features, but would not 'see past' objects.

    SAR works based on data collected over time as the antenna pair moves. The integration area becomes the new virtual antenna size and provides azimuth resolution. For best processing, you need to pass the object you are trying to resolve. For a vehicle, this means it would help is you were trying to image the buildings on the side of the road (esp if they are set far back), but it wouldn't help forward imaging appreciably. It also doesn't do well if things other than the antenna are moving.

    Antenna beam width, size, and frequency are proportionally linked. Automotive radar use of high frequency allows for small packaging. FWIW Magellan used a 12.7 cm wavelength (2.36 GHz vs automotive in the 24 to 77 GHz range).

    Agree, flooding would require baseline road height data (no way to know depth otherwise) Also need to handle loss of roads entirely like during the CA earthquake or the I-35W bridge collapse...
     
    • Like x 1
  5. scaesare

    scaesare Well-Known Member

    Joined:
    Mar 14, 2013
    Messages:
    7,324
    Location:
    NoVA
    I'd argue that throwing more I/O (sensor input), at a platform that doesn't have the capacity to "reason" doesn't necessarily work, and indeed can be counter productive.

    If it did, we'd have a brute-force platform with $50mil in sensors that drove perfectly, even if it was just a military demo. But we don't... we have platforms that indeed _DO_ have loads of sensor data, but they still run in to ditches.

    Without sufficient "reasoning skills" (AI capability) you are forced to try and reconcile that data with conditional set programming, and you simply can't cover all the possible scenarios. What's more, when you get conflicting input (which goes up exponentially with the # of sensors), the platform with insufficient AI is going to fail harder.

    If Lidar says there's an object there, and radar doesn't see it at all (because of the material type) and video cant pick it out against a similar background, what do you do?

    What if lidar sees it, but radar thinks it's much larger due to the concave shape of the object?

    Etc...

    It's the AI capability that allows for correctly evaluating these things.

    Not to say that sufficient sensor input is unnecessary.. it certainly is... but you can't just substitute more sensors for less AI.
     
    • Like x 1
  6. KarenRei

    KarenRei ᴉǝɹuǝɹɐʞ

    Joined:
    Jul 18, 2017
    Messages:
    6,050
    Location:
    Iceland
    You're moving through an environment full of beings that will be making decisions that can affect you. Making proper decisions concerning others requires a working Theory of Mind for those surrounding you. Adults driving big heavy objects. Children. Animals. Anything that can make decisions.

    The basics is just physics, but everything outside the base use-case involves logic. Hey, on this country road, should I be driving down the centre or staying in my lane? Well, I don't know, how well do I think that shoulder is going to hold up? How sure am I that there's not going to be a bad pothole just over this rise that's going to make me lose traction? If I go off the road, am I going to end up in grass or going down into a ravine? Aka, is whatever's on the side of the road going to kill me or not if I make a mistake? How do I know? That moss over there sure looks fluffy, but I know that what's underneath it is *lava*, and if I hit that, my car is going to turn into a flaming wreck. And what about those sheep grazing up there? Oh, that's okay, that's two ewes, I should only show moderate caution. And that's two rams, okay, no problem. And a lamb and a ewe on the same side, no problem. Uh oh, lamb and a ewe on opposite sides, I should slow down, because I know that lambs like to run to their mother when cars approach! Or maybe it's winter. Oh fun, it's time for a two-hour game of "strain your brain trying to figure out how slippery the road surface ahead of you looks!", something that's hard for even humans to estimate correctly.

    The basics is just physics. Want to get beyond the basics in order to deal with randomness and edge cases? Your task becomes far more difficult.

    And as a consequence, off-perpenduclar reflections appear brighter, and as a consequence, brightness is a measure of roughness, all other factors (surface angle, surface material, etc) being the same.

    Of course SAR is inapplicable here, which is why I quickly dismissed it and pointed out that a standard phased array antenna can do the exact same thing. Your car isn't going hundreds or thousands of meters per second. ;) SAR was only brought up because the radar maps we get from space are generally SAR, because unlike cars, they are moving thousands of meters per second. ;) And it makes more sense for a spacecraft to take advantage of that fact than to include a physically-large (and consequently heavy) antenna.
     
    • Like x 2
  7. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,998
    Location:
    Michigan
    My understanding:
    Tesla is swapping out the entire AP module, not the MXM daughterboard.
    The Tesla processor is dual core (fully redundant) and runs the NN 10x as fast. Speed boost due to reduction of memory moves and the logic optimizations you mentioned.
    Same basic module for all three vehicle types (housing likely differs, chips can remain the same)

    Tesla claims to have ‘world’s most advanced computer for autonomous driving’ with Autopilot 3.0 update coming next year

    Already tested with roll out in 4-6 months.

    Tesla (TSLA) Q2 2018 Earnings Conference Call Transcript

     
    • Informative x 2
    • Like x 1
  8. Fact Checking

    Fact Checking Active Member

    Joined:
    Aug 3, 2018
    Messages:
    4,373
    Location:
    Vienna
    What are you basing the 'dual core' part on - is it based on what Elon mentioned:

    But it's an incredible job by Pete and his team to create this, the world's most advanced computer designed specifically for autonomous operation. And as a rough sort of whereas the current NVIDIA's hardware can do 200 frames a second, this is able to do over 2,000 frames a second and with full redundancy and fail-over. So, it's an amazing design and we're going to be looking to increase the size of our chip team and our investment in that as quickly as possible. I think we have some of the best aces in the world, but I think we want to build on that even more.​

    ?

    If it's "dual core" I'm wondering how they'd do fail-over in such a design. SpaceX's software fail-over uses three chips, that way it's a two chip majority vote for what the computing result is. But that is 3x the chip hardware overhead.

    But yes, 'full redundancy' suggests at minimum duplicated cores.
     
    • Informative x 1
  9. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,998
    Location:
    Michigan
    A standard phased array cannot equal a SAR unless the array approaches the size of the virtual antenna. Phased arrays allow for beam steering and defocusing, but they cannot focus the beam (azimuth resolution) more than their physical size allows.

    Regarding wildlife handling. My thought is it comes down to how aggressive one wants the car to be. A sheep (or any animal) has a top speed. Based on that, the car, with rudimentary target recognition, can determine how far the animal can go in a time span. The car can go at a speed such that it likely will not strike the animal, that would be effective, but annoying to we human drivers.
    It also does not rule out collisions due to animal striking the vehicle (such as deer taking out side doors).

    The logic path you spell out is a probability set based on further assumptions (that's may be a lamb and ewe, but is it her lamb?). If I go slow, is it more likely to charge? Can one apply any logic to small children?

    In boating, the fail safe approach in fog is to navigate at a speed such that the boat can stop in one half the visibility. Something similar would be the safest car, but then take the case where the car is in a residential neighborhood. At any point a child could jump out from behind a parked car. Or could be hiding under a pile of leaves. Rare, but it has happened.

    On the upside for AP cars, they are always looking everywhere, so it removes the case that one looks away from the road the moment before an object enters the collision space. A 3 MPH car is almost intrinsically safe, but no one would use that. People trade off risk of impact for reduction in travel time. I'd hate to be on the committee that decides what a safe probability level is.
     
    • Informative x 2
  10. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,998
    Location:
    Michigan
    Yeah, full redundancy is the basis for dual core. You are right in that it could be 3. SpaceX designs allow for two failures before it approaches loss of mission. For dual soft failures in the same unit, it might be difficult to determine the errant part of the two remaining. After the first failure, there is no majority. However, having a test data set on-module could allow quick vetting of performance.

    Fortunately, cars are stable when unpowered so once the two cores disagree, it can enter a fail safe mode if it can't determine who's wrong. (Not that throwing on the hazzards and limping to the side or stopping in place us great, but that's what drivers do now, DFMEA and all that)
     
    • Informative x 1
  11. KarenRei

    KarenRei ᴉǝɹuǝɹɐʞ

    Joined:
    Jul 18, 2017
    Messages:
    6,050
    Location:
    Iceland
    #11 KarenRei, Aug 31, 2018
    Last edited: Aug 31, 2018
    You don't have to equal it. You're not imaging the surface of a planet hundreds of kilometers below you. ;)

    Resolution is not particularly important for this task. 15cm at 100m would be great. Even 35cm at 25m would be useful.
     
  12. Fact Checking

    Fact Checking Active Member

    Joined:
    Aug 3, 2018
    Messages:
    4,373
    Location:
    Vienna
    Indeed, even just the ability to detect failure (disagreement between cores) is an important metric to hardware reliability.

    It doesn't have to be able to determine who was wrong - displaying an alert and restarting both would be useful already.
     
    • Like x 1
  13. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,998
    Location:
    Michigan
    #13 mongo, Aug 31, 2018
    Last edited: Aug 31, 2018
    I think you may be misunderstanding the antenna sizes needed and how the system works. In a chirped SAR system, 15cm or 25cm resolution requires two things: sufficient pulse bandwidth for the range resolution and sufficient virtual antenna size for the azimuth resolution. The distance to target does not impact either of those. Distance to target does matters in terms of setting your receive range gate and the motion compensation needed. The beam itself also spreads out which allows you a larger potential integration path (object of interest must be illuminated) when imaging a planet, but airborne and close in ground systems use SAR just fine. (also need to balance beam width with transmit power/ receiver sensitivity) If there is height variation, a long look might not even be the best approach due to shadowing.

    For a steered phased array, there is only the physical size, which determines the beam width which determines the azimuth resolution (which get worse with distance). Single emitter multiple receiver setups help, but I think resolution is still limited to their physical baseline (VLA for instance).

    Good slide deck on the SAR for any interested parties:
    Radar 2009 a 18 synthetic aperture radar

    Edit: Regarding satellite imaging, yes, to achieve a high resolution satellite image with a physical antenna would be highly impractical due to the very narrow beam needed due to distance. For close in work, the spread is much less, but 15cm at 100m is less than a 0.1 degree beam width, 25cm at 25m is roughly 0.6 degrees.
     
    • Helpful x 1
  14. KarenRei

    KarenRei ᴉǝɹuǝɹɐʞ

    Joined:
    Jul 18, 2017
    Messages:
    6,050
    Location:
    Iceland
    Could you kindly explain why you're still talking about SAR here? I only presented SAR images because that's what's used for radar imaging from space due to how fast spacecraft travel. You can create such a reflection map with any radar technology; SAR is just most appropriate for spacecraft, given their operating constraints.

    This is false. Angular resolution depends on the antenna's beamwidth. Physical resolution depends on angular resolution times distance. Or to put it another way: there's a reason that those satellite radar maps aren't imaging objects the size of pennies. Distance absolutely matters.

    Aka, beamwidth. Which yields angular resolution. Which, times distance, yields physical resolution.

    Exactly my point when I wrote, and I quote, "Resolution is not particularly important for this task". It's not a limiting factor, as you were making it out to be when bringing up the smaller physical antenna size of a carborne phased array vs. the virtual size of a SAR. A small carborne phased array is more than sufficient to measure returns from the road at sufficient resolution at sufficient distance to build up a road roughness / materials map, and react to that. You don't need a gigantic virtual aperture. You're not imaging a planet from hundreds of kilometers away.

    (It's also worth pointing out that aperture size corresponds linearly with wavelength. Of course, on Earth you can't just pick any frequency you want to broadcast in, but....)
     
  15. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,998
    Location:
    Michigan
    Because you were referencing satellites which use SAR. (and I like SAR)

    from slide 34, "The [cross range] resolution of a focused SAR is independent of range and the wavelength and depends solely on the dimension D of the real antenna" I realized I mis-attributed this quality to standard radar also, hence my post edit.

    You also wrote:
    hence my quick calculation.
    If you are talking about roughness, that is a different issue than object resolution. However, if you did want 25cm object resolution at 25 m, then at 77 Ghz, you would need a 0.6 meter dish for 6 db down. At 24 GHz, it would be 2 meters.

    True that: ;)
     
  16. KarenRei

    KarenRei ᴉǝɹuǝɹɐʞ

    Joined:
    Jul 18, 2017
    Messages:
    6,050
    Location:
    Iceland
    #16 KarenRei, Aug 31, 2018
    Last edited: Aug 31, 2018
    I said 35cm at 25m (I think being able to steer about a foot to the side in about a second of travel is reasonable, even in slippery conditions, don't you?), but let's go with 25cm, aka 0,25m at 25m distance, a 100:1 ratio, or 0,5229 degrees, or 0,01 radians. Beamwidth is wavelength times beamwidth factor over aperture, and we want to solve for aperture, so aperture = wavelength * beamwidth factor / 0,01. Substituting 77GHz, aka 0,0038934m for wavelength and using a beamwidth factor of 1,2 we get a 0,47 meter aperture. 35cm would be 0,33m, aka about a foot.

    That said, you don't have to have a "dish" of that size. The optimal implementation if you want radar resolution is two antennas separated from each other, one on the left side of the car and the other on the right. With that you could have an effective aperture of 1,8m.
     
  17. surrealize

    surrealize New Member

    Joined:
    Jun 5, 2018
    Messages:
    1
    Location:
    Emeryville, CA
    Re: radar, I was impressed by the radar resolution in this image:
    automotive high res radar vs lidar.png

    from this article from last November about R&D work on imaging radars for autonomous driving:
    Here Comes High-Res Car Radar
     
    • Informative x 2
    • Like x 1
  18. Causalien

    Causalien "The" 8 ball Oracle

    Joined:
    Nov 19, 2012
    Messages:
    2,461
    Location:
    Pothead's Republic of Canukstan (PRC)
    Ya, that's one benefit of using your own designed asic and pcb. Instead of having pcie to send info in serial, you can just have a massive parallel bus with as many bits as you want to send info within 2 clock cycle.
     
  19. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,998
    Location:
    Michigan
    Totes, especially if you can distribute the intermediate memory on chip between the pre-placed multiply/accumulate, soft max , sigmoid logic units.
    Camera feed (serial) in, commands out.
    Could be really low (data) pin count.
     
  20. Causalien

    Causalien "The" 8 ball Oracle

    Joined:
    Nov 19, 2012
    Messages:
    2,461
    Location:
    Pothead's Republic of Canukstan (PRC)
    It'll suck to be the pcb board designer troubleshooting timing issues :)
     

Share This Page

  • About Us

    Formed in 2006, Tesla Motors Club (TMC) was the first independent online Tesla community. Today it remains the largest and most dynamic community of Tesla enthusiasts. Learn more.
  • Do you value your experience at TMC? Consider becoming a Supporting Member of Tesla Motors Club. As a thank you for your contribution, you'll get nearly no ads in the Community and Groups sections. Additional perks are available depending on the level of contribution. Please visit the Account Upgrades page for more details.


    SUPPORT TMC