Tesla Autonomous Driving H/W

[scaesare]

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.

We are talking a minimum factor of 10 in power consumption and cooling required.

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.

What are you talking about- this is probably most absurd thing I have heard.

Do you even realize the amount of data bandwidth is needed? Embedded memory isn’t cheap. Certainly more expensive
Than external memory.

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.

I got to laugh on this one too.

Just because nVidia uses terms GPU. Doesn't mean its exclusive.

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.

Do you even understand what a custom ASIC is? Nothing is build from Scratch anymore. Tesla most likely will be use a version of ARM core for the main CPU. Probably some sort of licensed GPU as well.

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.

Even Apple own processors are ARM based. Those are not free.

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

Yes. So I understand how ARM license works

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

Son- please go read up what lidar does and why it is needed. You are embarrassing yourself.

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

I work in this field. You have zero clue on AD.

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.
​

 

[lunitiks]

Gotta love invisible ink

 

[scaesare]

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

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.

 

[mongo]

Does AI have the reasoning skill of a human?
Will it any time in the near future?
No? Then it needs better senses to compensate for its failings.

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).

That said, I don't think LIDAR is key. I fully agree with Elon - if you're going to have a transmit-and-receive system, why would you double up on the same spectrum as your cameras, and face that spectrum's same limitations? Radar is the way to go. Sure, the radar spectrum has its own weaknesses, but they're different weaknesses from the optical spectrum.

That said, I don't see them yet using radar to its full potential. As far as anyone can tell, they're just using it as an object finder / distance to object calculator, like LIDAR. But radar has the potential to be so much more than that. I wrote an article on M3OC a while back about this, but the short version: have you ever seen a satellite radar map (generally made using SAR, but any radar can do it, incl. small vehicle-borne phased arrays)?

That's Baltis Vallis (Venus), the longest riverbed in the solar system. Here you're not focused on the timing of the echos, but rather brightness (signal intensity). What determines the brightness? It's partially the material and the angle, but beyond that, you're looking at the roughness of the surface on the scale of the wavelength. Where you see white, that's rough areas, while darkness is flat areas. By changing the wavelength, you can probe the surface roughness on different scales - anything from texture the size of grains of sand or smaller, to texture the size of potholes or larger.

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.

I imagine the radar they're using currently is locked into a single band. But even that could potentially be useful information. And there's always potential for more capable radars in the future. This sort of "information from senses that we don't have" gives autonomous vehicles the potential in the future to compensate for their lack of human-level reasoning skills.

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).

Another example of how "information beyond our senses" can mitigate problems is the case of flooding. You don't want your car stopping on the freeway due to a puddle, but you also don't want it just driving into deep water and drowning your car (and potentially killing you).

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...

 

[scaesare]

Does AI have the reasoning skill of a human?
Will it any time in the near future?
No? Then it needs better senses to compensate for its failings.

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.

 

[KarenRei]

Autonomous driving doesn't really reasoning to work/ be safe

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.

Rough surfaces give a more omni-directional return

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.

SAR works based on data collected over time as the antenna pair moves

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.

 

[mongo]

After KrispyKreme's technical assault, I decided I Need to understand the electronics that handles the ai. So I started the search 2 days ago and finally found the thread with AP circuit board teardown and subsequent discussion today. Then took a deep dive into Nvidia's Parker (TA795SA-A2) general purpose gpu architecture and also the discreet Pascal GPU (GP106-510-KC) used. It was a deep dive, but I came out of it understanding that it is possible for Tesla to drop in a chip and enable Full Self driving. I will try to be as non technical and easy to understand as possible.

There were many experts chiming in both on this forum (for the sake of the person doing the teardown, will not link to it) and on Reddit's thread. However, not many have all the specialties required to see the full picture. I happen to have been an ASIC designer, firmware engineer, Machine Vision Application Engineer and in a jam, my managers have forced me to solder and troubleshoot PCBs so I know a bit of that. The only discipline I do not have is chip layout/routing and PCB layout/routing. (yes I jump ship every 3 years, jack of all trade and not an expert in any)

Here's my guess of how they use the hardware from looking at just the architecture.

Nvidia's Parker general purpose GPU is probably what Tesla intends to replace. Contained within is a multipurpose ARM Cpu and a small piece of its Discreet GPU. These GPUs have a bunch of units that process pixels in parallel and shove them into what they call Tensor units. Tensors units are programmable units that can be changed for other operations. I believe that these are used as the "Neurons" in neural network for decision making.

Some GPU parts not needed for AI

Nvidia, probably in a rush to bring a product to market, did not really design a GPU specifically for Neural Network, instead they brought their gaming discreet gpu and stuffed it with some control logics and called it a day. The Processing part of their GPU has a lot of waste. Things specifically made for gaming and displaying images can probably be taken out completely. Also all the calculations can probably be narrowed down to int8 once the neural network is sufficiently trained and the results can be locked down. For simple pre-processing, drawing contours and recognizing an object in an image. a 8 bit black and white image will suffice and TSLA uses the red spectrum to do that. Currently, many calculations passes through 16 bit floating points and 32 bit floating points (most likely necessary for training the neural network). These are very expensive operations and take up a lot of spaces.

AP 2.0 Model S, AP 2.0 Model 3, AP 2.5 Model 3

Model S tear down shows that it has 1 Parker general purpose and 1 discreet gpu. Model 3 tear down from Munroe shows 1 Parker general purpose and 2 discreet gpu. So there's some upgrade there. Then there's a potential AP 2.5 hardware showing up where extra connectors to a potential new board are present. In the future, there may be two board linked together to perform Full Self driving. My own guess is that it will evolve to two of the Nvidia board in parrallel for a while before Tesla have finished designing and testing their own chip. Both of the chips seems drop in replaceable as they both seems to sit on MXM. My guess is that there's a strong chance TSLA replaces both the General purpose and Discreet GPU.

I personally haven't done any calculation on how many chips is necessary to achieve full self driving. Potentially AP 2.0 Model S might have some trouble, but Tesla can just tapeout a chip with 7nm process for the older Model S which effectively doubles the amount of operations it can do. But this probably won't be possible if we stick to Nvidia solutions for the interim.

COST

However, the question is, is it worth it. Nvidia is forging ahead with their new plex 2.0 platform. More power consumption more heat generated, but it'd save TSLA the R&D cost as well as the design and tape out cost + equipment for analysis. I wouldn't be surprised that it will cost 1 mil or 2 to tapeout the first batch of chip (plus ~3$ a chip) and the chip design team of around 20 person x 100k salary (probalby 200k + if TSLA is using the best) x 2 years of time + equipment. On top of that, Taping out is a very rigid process and cannot be modified too much on the fly. There's only so many Engineering Change order you can do before the mass production.

What is so bad with sticking with NVidia and just increase cooling and power consumption, which was the reason why KrispyKreme was attacking Tsla's automated driving electronics. That reasoning is lost to me.

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

Elon Musk -- Chairman, Product Architect, and CEO

Yeah, it may be worth articulating some of the details, design principles that like explain why the Tesla AI chip, or AI computer, essentially, for the car is able to achieve an order of magnitude better processing than anything else that exists. Yeah.

Peter Bannon -- Director of Silicon Engineering

Sure. So, like three years ago when I joined Tesla, we did a survey of all of the solutions that were out there for running neural networks, including GPUs. We went and talked to other people like at ARM that were building embedded solutions for running neural networks. And pretty much everywhere we looked, if somebody had a hammer, whether it was a CPU or a GPU or whatever, they were adding something to accelerate neural networks but nobody was doing a bottoms-up design from scratch, which is what we elected to do.

We had the benefit of having the insight into seeing what Tesla's neural networks looked like back then and having projections of what they would look like into the future, and we were able to leverage all of that knowledge and our willingness to totally commit to that style of computing to previous design that's dramatically more efficient and has dramatically more performance than what you can buy today.

Elon Musk -- Chairman, Product Architect, and CEO

Cool. Thanks. Yeah, I mean, essentially the key is to be able to run the neural net at a bare metal level so that it's especially doing the calculations in the circuits itself and not in some sort of emulation mode which is how a GPU or a CPU would operate. So, you want to do basically a massive amount of localized matrix multiplication with the memory right there.

So, it's a huge number of very simple complications with the memory needed to store the results of those complications right next to the circuits that are doing the matrix calculations. And the net effect is an order of magnitude improvement in the frames per second. Our current hardware, which ... I'm a big fan of NVIDIA, they do great stuff but using a GPU, fundamentally it's an emulation mode, and then you also get choked on the bus. So, the transfer between the GPU and the CPU ends up being one that constrains the system.

So, the net effect is we're able to, with the Tesla computer ... and we've been like semi-stealth mode basically for the last two to three years on this but I think it's probably time to let the cat out of the bag because the cat's going to come out of the bag anyway 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 there's a rough sort of [Inaudible] whereas the current [Inaudible] 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 [Inaudible] in the world but I think we want to build on that even more.

And it costs the same as our current hardware and we anticipate that this would have to be replaced, this replacement, which is why I made it easy to switch out the computer, and that's all that needs to be done. If we take out one computer, plug in the next. That's it. All the connectors are compatible and you get an order of magnitude, more processing and you can run all the cameras at primary full resolution with the complex neural net.

So it's super kick-ass. Thank you for doing that.

Peter Bannon -- Director of Silicon Engineering

You're welcome.

 

[Fact Checking]

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)

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.

 

[mongo]

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.

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.

 

[mongo]

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.

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)

 

[KarenRei]

A standard phased array cannot equal a SAR unless the array approaches the size of the virtual antenna

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.

 

[Fact Checking]

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)

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.

 

[mongo]

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 25cm at 25m would be useful.

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.

 

[KarenRei]

I think you may be misunderstanding the antenna sizes needed and how the system works. In a chirped SAR system...

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.

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

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.

The beam itself also spreads

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

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.

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....)

 

[mongo]

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.

Because you were referencing satellites which use SAR. (and I like SAR)

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.

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.

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.

You also wrote:

15cm at 100m would be great. Even 25cm at 25m would be useful.

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.

(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....)

True that:

Antenna beam width, size, and frequency are proportionally linked.

 

[KarenRei]

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.

 

[surrealize]

Re: radar, I was impressed by the radar resolution in this image:

from this article from last November about R&D work on imaging radars for autonomous driving:
Here Comes High-Res Car Radar

 

[Causalien]

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

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.

 

[mongo]

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.

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.

 

[Causalien]

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.

It'll suck to be the pcb board designer troubleshooting timing issues