Tesla's Full Self-Driving Computer is Now in All New Cars and a Next-Gen Chip is Already 'Halfway Done' 173
The Tesla computer, a new custom chip designed to enable full self-driving capabilities, is now in all new Model 3, X and S vehicles, CEO Elon Musk said during the company's Autonomy Day. From a report: Tesla switched over from Nvidia's Drive platform to its own custom chip for the Model S and X about a month ago and for the Model 3 about 10 days ago, Musk said. "All cars being produced all have the hardware necessary -- computer and otherwise -- for full self-driving," Musk said. "All you need to do is improve the software." Work is also already underway on a next-generation chip, Musk added. The design of this current chip was completed "maybe one and half, two years ago." Tesla is now about halfway through the design of the next-generation chip. Musk wanted to focus the talk on the current chip, but he later added that the next-generation one would be "three times better" than the current system and was about two years away.
ANY DAY NOW (Score:1)
Flying cars to Mars.
Fo sho.
Re: (Score:1)
Oops I got mud on the 7 camera's, 1 forward facing radar and 12 ultra sonic sensors.
FTFY
Re: Retrofit? (Score:1)
Oops. A five gallon pail of driveway tar fell out of my truck bed, and covered the entire front end of your wonder car. Oops.
Re: (Score:1)
Re: (Score:2)
Re:Retrofit? (Score:5, Informative)
So how much will it cost Tesla to upgrade the millions of cars they've already produced and whose owners have already paid for FSD? $Billions? They're broke and have not reserved any $$$ for that in their financials.
They haven't sold millions of cars yet - as of December 2018, they had sold approx 500 000 cars [wikipedia.org]. Some of these no longer exist, and a healthy part of them probably did not come with FSD paid for - given that there was no actual features part of the option and very few thought Tesla would be able to actually do full self drive within the lifespan of the cars, skipping it was a sensible option.
Of course, getting a hardware upgrade is nice even though I'm still pretty sure they're not going to be able to do FSD for many, many years. The hardware isn't capable enough (no LIDAR), the software is nowhere close to ready even for best case scenarios (good weather, good markings) and after that there's a long regulatory road ahead.
Re: (Score:2)
The 2.5 hardware and current software is amazingly good at seeing the lane markers, even when the road only has temporary markings.
Re: Retrofit? (Score:2)
very few thought Tesla would be able to actually do full self drive within the lifespan of the cars
And very few* still do.
*Don't get me wrong; our man Elon has no bigger fan; that having been said, as a 31year 'IT Nerd' and professional driver for even longer... I see through the self-driving craze.
Re: Retrofit? (Score:2)
Re: Elon Musk is a billionaire (Score:1)
Musk is a plutocrat. They have a tendency to suffer difficulties the rest of us can only imagine during major civil disturbances.
Re: (Score:2)
One of the defining characteristics of a plutocrat is a person who is in it only for themselves and actively work to increase poverty. Think of all the people who would fit that description - Elon is not one of them
LIDAR is silly. (Score:5, Insightful)
Re:LIDAR is silly. (Score:5, Interesting)
The problem is that the claimed AI isn't an AI. It's an interference engine that's incapable of following a lane if there's leaves on the roadway. There's no real intelligence behind these AI systems, they have the smarts of a cockroach. They can see patterns and follow algorithms based on those patterns but any disruption outside the defined parameters and they are dead in the water.
A perfect example of this was the guy on autopilot that ran into an attenuator at 70mph, The inside stripe on a diverging ramp had faded out of view to the cameras on the car followed the other yellow stripe right into the attenuator at full speed. The software couldn't even do the basic deductive loop of the stripe on the left not matching the stripe on the right and looking forward on the roadway to see what to do. Something even a 3 year old can do. People give these systems far to much credit for the capabilities they actually have.
AI isn't intelligent, it's not even close to what the majority of people think of as AI. Real AI is decades away at best. These systems are at best advanced cruise control systems with the ability to do certain tasks under idealized conditions and they fall on their ass in situations where human drivers don't even blink. Such as leaves on the roadway.
Re: (Score:3)
The problem is that the claimed AI isn't an AI. It's an interference engine that's incapable of following a lane if there's leaves on the roadway. There's no real intelligence behind these AI systems, they have the smarts of a cockroach.
I think you're insulting cockroaches by insinuating they are as stupid as the AI systems...
Re: (Score:2)
In general, I think there is more one can do than just add more training data. You can add more sensors/information. You can add redundant learning systems based on the new information. You can change the architecture in fundamental ways. However, that doesn't address your main point on how one validates the system.
Regression testing is probably OK, since yo
Re: The same applies to humans. (Score:1)
Yes. The human brain is just a mechanism that we nearly have figured out.
Right.
Re:LIDAR is silly. (Score:4, Interesting)
https://techcrunch.com/2019/04... [techcrunch.com]
“Lidar is a fool’s errand,” Elon Musk said. “Anyone relying on lidar is doomed. Doomed! [They are] expensive sensors that are unnecessary. It’s like having a whole bunch of expensive appendices. Like, one appendix is bad, well now you have a whole bunch of them, it’s ridiculous, you’ll see.”
Re:LIDAR is silly. (Score:4, Informative)
The kind of image recognition that the human brain can do is way beyond what the most powerful computers are capable of, even given unlimited time to process the data. For example, humans can very quickly recognize the 3D shape of objects, separate them from the background, and understand the layout of the environment around the car.
Tesla can't even recognize the sides of the road unless there are reasonably clear markings. They are still trying to get basic image recognition to work when to do things like navigating around road works or a car park they will need to be able to understand fairly arbitrary 3D shapes by comparing sequential images, map them into a model of the local environment and keep track of them as they move or become obscured temporarily.
The reason everyone else uses lidar is that those problems have not been solved even with unlimited computing resources, let alone in a fraction of a second on a relatively low power system. If Tesla ever does find a viable solution it then still has to do all the other work actually making the thing drive safely like Waymo did.
Keep in mind that Tesla promised far more than Waymo have delivered so far, and a full demo in 2017. Take anything they say now with a metric tonne of salt.
Re: (Score:3)
separate them from the background, and understand the layout of the environment around the car.
It's important to note your post is a mixed bag of true and false. The human's brain excels in the understanding of what it's seeing, that "recognition" part. However computers can trivially separate objects from backgrounds just fine and they are better at understanding what's around the car than humans are.
You say Tesla can't recognise the sides of roads without markings? That is surprising given I've seen a video a Tesla driving completely hands off following a completely snowed under road with the only
Re: (Score:2)
Do you have a link to this video?
It seems extremely improbable, given that the cameras would have had to detect the white-on-white bump in the snow, with enough confidence for AP to keep going.
Every video of AP in snow either shows a failure or it following a car in front, or the road is visible enough for it to follow.
The lack of road markings issue is noticeable in videos of unmarked roads, and particularly where road markings are temporarily obscured or removed for some reason (e.g. at a fork in some cou
Re: (Score:2)
But human eyes, even with the astonishingly masmart processing the human brain can do, are demonstrably NOT good enough - how the hell can 1.25 million deaths a year be considered "good enough"? What about when it's dark? Foggy? Blowing snow? Glaring low sun?
Sometimes LIDAR is going to be better, sometimes optical is going to be better. Having both (and maybe other detectors too, infrared for instance) give the possibility of being much better in most situations, if not all. Add in detectors such as skid de
Re: (Score:3)
But human eyes, even with the astonishingly masmart processing the human brain can do, are demonstrably NOT good enough - how the hell can 1.25 million deaths a year be considered "good enough"?
Most of these 1.25 million deaths are not due to the driver's eyes not being good enough.
Re: (Score:2)
As I said, eyesight is one thing that can certainly be improved by using multiple cameras. My point was that if you add that in to the other sensors and controls the self-driving car will have that are are either unavailable to or simply faster than human drivers, the advantages start to really add up.
Computers already better than humans at this (Score:2)
The human brain is adept at discarding useless information
The problem is, the brain can change the definition of "useless" to include driving related stuff, when attention is diverted elsewhere - like great scenery. I cannot count how many times I have seen drivers crossing middle lines of roads because of food, scenery, conversation with occupants... I was a passenger in one guys car who mast have had a two foot variance in the distance between him and a large concrete median to the left of the car. The
Re: (Score:2)
That's why you use a neural net.
Re: (Score:2)
To see this for yourself, generate some sine of cosine data between 0 and 2pi. Than a neural network on this data. Then feed in 3pi and see if you get the right answer.
This is meaningless. Just because you are familiar with sine and cosine functions does not mean that this is the only, or optimal way to extrapolate the data.
Re: Retrofit? (Score:1)
Also: the new chip.
Where have we heard that before? You mean last month's new chip? Does it reall upgrade over wireless? What if smoke and mirrors gets in the signal path?
Re: (Score:3)
Re: (Score:2)
They've been bilking people out of $5k for this "feature" that has never come for the last 3 years
My spouse bought her Tesla in 2015, and paid for FSD. So 4 years, not 3.
Re: Retrofit? (Score:2)
The new processor is cheaper and uses less power than the old. Simple plug in replacement. All the sensors are already in place.
Hardware and software... (Score:5, Insightful)
"All cars being produced all have the hardware necessary -- computer and otherwise -- for full self-driving," [...] Work is also already underway on a next-generation chip
Either the current hardware is not sufficient, or the next generation is unnecessary.
Oh, and by the way, "All you need to do is improve the software" is an incredibly oversimplification of the problem of autonomous driving.
Re: (Score:1)
Musk also said that Tesla's had full self-driving hardware back in 2016. He was lying then too.
https://www.cnet.com/roadshow/news/elon-musk-says-all-tesla-models-will-have-self-driving-hardware/
Re:Hardware and software... (Score:5, Insightful)
Either the current hardware is not sufficient, or the next generation is unnecessary.
The next generation hardware could provide the same features with much lower power consumption, leaving more energy for driving. The batteries still have limited capacity.
Re: Hardware and software... (Score:3, Insightful)
The percentage of power devoted to autonomous driving is very small, simply because the chips would overheat and die if they required more than 100-200W. That's way less than 1 HP, so power consumption, while relevant, is not a limiting factor. Features are.
Re: (Score:2)
Re: (Score:2)
Do you have a reference for this? I'm honestly curious - Last I heard autonomy stacks were pulling around 5kW, causing problems for range of all-electric autonomous vehicles. 5kW is a significant increase over highway cruise power of 15-20kW...
A 95% reduction in power requirement for autonomy hardware is massive.
NOTE: this is total, I suspect that a single chip is of course much lower.
Re: (Score:2)
LIDAR based systems are power hogs. Tesla's neural net chips don't take much power. The neural net board is a drop in replacement for the current MCU and power consumption is the same.
Re: Hardware and software... (Score:1)
Can I interest you in a used Altivec Unit? It also has SCSI.
Seriously. Tesla is going to starve Apple of their stupidest, most belligerent customers.
Re: (Score:2)
The beer-cooled computer does not harm the ozone layer. -- John M. Ford, a.k.a. Dr. Mike
Re: (Score:2)
I watched the live stream this afternoon, and IIRC they said their custom CPU draws about 230W, which is probably two or three percent of the overall energy budget for the car. So yes, it's significant enough that you'd want to address it. Also, if you can increase the computing power, it may result in a safer car, which could be even more valuable.
Re: (Score:2)
Re: (Score:2)
I think you confused "zero" with "millions" (Score:2)
Tesla logged ZERO miles of autonomous testing in 2018. ZERO.
Part of testing is data collection as to what sensors see from a wide variety of sensors.
Who is better off, Waymo which maybe has one car that has ever seen snow - or Tesla with hundreds of thousands of cars operating in all kinds of climate?
Tesla cars in consumer hands have already logged millions of self-driving miles, even if in most cases drivers were also monitoring - and Tesla knows when drivers had to correct, further improving training...
Yo
Re:I think you confused "zero" with "millions" (Score:4, Informative)
It's a myth that Tesla is racking up millions and millions of miles of training data. They are not streaming data from all their cars in real time. It would cost a fortune in mobile data, aside from anything else.
At most they are identifying areas where Autopilot consistently does something very different to what the human driver does, or where it consistently craps out and the driver has to take over.
Waymo is very, very far ahead because they have better sensors, which means they have a better map of the environment, which means they can concentrate on the actual process of driving and not struggling to just understand what the cameras are seeing. Tesla has yet to demonstrate any kind of 3D environment mapping, which is absolutely essential to self driving.
That is not right. (Score:2)
Of course they are not streaming data from everyone driving in real time.
However they are pulling car logs from cars brought into service, and of course they get feedback from drivers as to performance.
Also check out things Tesla is sending back, like locations where Autopilot is disengaged - they can try running other cars in those areas to see what issues might ben encountered and fix the problems.
They could also be sending a lot more back on recognition performance when you arrive home and are connected
Re: (Score:2)
3D mapping and stereo vision is not the same thing.
With 3D mapping people mean, looking front, left, right and back continuously and having a 3D model of your suroundings, not simply a set of stereo views.
P.S. - confirmation of 3D processing (Score:3)
Tesla has yet to demonstrate any kind of 3D environment mapping,
If you'd actually watched the YouTube presentation (long I know but still), you'd see that the vision system is building a 3D environment internally.
But again, that was obviously the case by the fact the system A) works, and B) has more than one camera for each facing...
P.S. - they are getting data from all cars... (Score:3)
They are not streaming data from all their cars in real time
After watching the video I understand better how this works - although all cars are not streaming, they can ask the entire Tesla fleet for images or video during driving - one example they gave is asking the fleet for images with bikes on the backs of cars. Or images with road debris. Or cars suddenly cutting off your car.
The video is actually very interesting and technical (the part I'm talking about is around 2 hours in).
Re: (Score:2)
Waymo is very, very far ahead because they have better sensors,
Which makes them also far behind, because they are now stuck to using expensive LIDAR forever, or having to redo all the work with optical cameras.
Re: (Score:3)
The plan is to get the cost and size of lidar down. There are many companies working on that, with solid state lidar looking most promising.
Tesla thought they could get ahead by just using cameras, but because that made the task so much more difficult it looks like lidar may overtake them anyway if any of the current exhibited prototypes work out. Certainly for commercial vehicles they only need to get the cost down, size isn't that important.
Re: (Score:1)
Tesla has yet to demonstrate any kind of 3D environment mapping, which is absolutely essential to self driving.
Explained here: https://youtu.be/Ucp0TTmvqOE?t... [youtu.be]
They have a 3D reconstruction of what the camera sees, and it's pretty good, but not as good as LIDAR.
Re: (Score:2)
which means they can concentrate on the actual process of driving and not struggling to just understand what the cameras are seeing.
That is no struggle at all, that is actually pretty simple.
What current cars lack is a complete 3D surrounding sensoring, that could be done with cameras, too. But research is focusing on LIDAR.
More proof Tesla is far ahead (Score:2)
Number of people killed by Tesla: 4
That too is test results Tesla uses to make newer cars better, no matter how grim a fact it it may be. In those cases Tesla gathered even more data than from normal drivers.
That absolutely proves Tesla has more than zero testing miles, and proves just how far ahead Tesla is. The fact that Waymo has made no serious errors means they are testing a lot slower than Tesla, and learning less.
I thank these early drivers for the sacrifice they made to improve future cars, thoug
Re: (Score:2)
Number of people killed by Waymo: 0
Number of people killed by Tesla: 4
Number of people killed by General Motors: 1.2 Million
Re: (Score:2)
Number of people killed by Waymo: 0 Number of people killed by Tesla: 4
Number of people killed by General Motors: (at least) 124
Fixed that for you.
Fixed that for you. https://www.caranddriver.com/n... [caranddriver.com]
Re: I think you confused "zero" with "millions" (Score:1)
Battery fires? Like the passenger roasted to death in the incident in Indianapolis. Nice dutch oven, that.
Re: (Score:2)
All of the AP 2 and 2.5 cars are analyzing and collecting road data all of the time (whether using AP or not) and sending this to Tesla. That's billions of miles of real world driving... orders of magnitude more than any other self driving company.
Re: Hardware and software... (Score:2)
...is an incredibly oversimplification of the problem of autonomous driving
It is not, however, an oversimplification of the problem(s) of brand awareness and shareholder value. ;)
That's a strange argument for Slashdot (Score:2)
Either the current hardware is not sufficient, or the next generation is unnecessary.
This is just like saying there is no reason you couldn't play Call Of Duty 3 on a SNES, or maybe more like saying because you could play a video gam on SNES, then the extra power of a PS4 means the SNES was never sufficient to play games on (false) or the next generation was unnesscary (also false).
More computer power is always more useful to be able to do more complex work. Just because something works, does not mean it c
Your argument already proved wrong (Score:3)
So outside of California, only the US south would be viable year round, New England only during summer, and PNW almost never.
Already have read reports from drivers in NYC using Teslas latest self-driving mode all the way form the highway into the city during a rainstorm, never having to intervene with the cars driving.
Snow is the only thing remaining that I could see fouling sensors. But since we already know how to keep headlights from being totally fouled by snow, I'm pretty sure that is possible for car
Re: (Score:1)
Thanks for letting me know, I thought maybe they had washers but didn't know they had heaters. I don't have a Tesla myself so was not aware of the exact details.
Re: (Score:2)
This is the same bullshit they have been feeding people for years. V2 was supposed to be capable of full self driving, coast to coast demo scheduled for 2017...
Now they have gone as far as redefining "full self driving, you can summon the car from the other side of the country" to "some driver aids, and you must keep your hands on the wheel at all times".
The demo was pathetic. They can recognize a few things. My phone could do that years ago.
Don't get me wrong, I still have a Model 3 reservation, but Tesla'
Re: Bought the wrong chair? (Score:1)
Is the stick mounted on your forehead? What's that all over your face?
'Halfway Done' (Score:1)
Tune in 6 months from now when Musk is sued for baiting people into buying stock with false company data.
I also suspect that Waymo, et. al would consider their efforts "halfway done".
All you need to do is improve the software
Oh, is that all? Just a couple of for loops I'm sure.
Re: (Score:3)
Ive heard this before (Score:2)
Elon and Tesla said this 2 years ago when I bought my P100D X (with FSD). Now the screen is so burned in it affects map use and they said they will get the screens âoethis summer.â
Iâ(TM)m losing faith.
I like to put it in puppy mode (Score:1)
Just hold down the three buttons and when the puppy icon pops up, select that, it allows the onboard nav to target all puppies and small children while "driving".
Full self driving mode works... (Score:3)
Re: (Score:1)
So does self-crashing.
Again? (Score:3)
"All cars being produced all have the hardware necessary -- computer and otherwise -- for full self-driving,"
Doesn't he make this same announcement about every 6-12 months?
Re: (Score:2)
If the car inputs and outputs and acutators / other mechanisms are in place, a fucking Z80 has the hardware necessary for full self-driving.
"It just needs the software".
No shit, sherlock. And that's what you don't have. And now that you've been mixing and matching models, chips, software versions, etc. people will never be quite sure whether their one is full self-driving or not and will assume they all are, even without the software upgrade (which I can't see actually happening, but that's another matter
Re: (Score:2)
a fucking Z80 has the hardware necessary for full self-driving.
No it has not. It has not even the address space to analyze a single still image, regardless how much time you give it.
Re: (Score:2)
Ehh . . . a Z-80 *could* drive the car with appropriate systems . . . but it's not going to be able to handle "side tasks" such as not running over pedestrians, turning at an appropriate time, . . .
hawk
All you need to do... (Score:2)
Will Tesla accept full liability for accidents? (Score:2)
EULA will not save them in criminal court! (Score:2)
EULA will not save them in criminal court!
Re: EULA will not save them in criminal court! (Score:1)
Criminal court is in session every day.
Re: (Score:2)
Since Tesla is, effectively, driving the vehicle, shouldn't Tesla be the liable entity?
No. The owner agrees to take over the liability. Or rather, the vehicle insurer of the owner.
Re: (Score:2)
I've not seen those words in any of Tesla's materials. All I see is how the car drives itself. So I'll ask again, will Tesla accept liability for their self-driving cars?
Right, right... (Score:4, Interesting)
If the obstacle to getting cars to self drive were that the problem required twice the power; or even an order of magnitude or two more power; than can reasonably be stuffed into a production car improvements to chips for the purpose would obviously be useful; but we would already have fully functional tech demo versions: few people would be interested in having the entire trunk and back seats occupied by a rack worth of compute gear that costs as much or more than the car carrying it and draws about as much power as the engine; but the fact that it's not going to be a consumer success yet wouldn't stop your R&D or PR people from brute forcing the compute requirements and showing off what will become viable after a few rounds of process improvement.
That is not obviously happening; and the people who are the closest to doing it are the ones loading up on extra sensors to supplement the cameras.
What is 3 times better. (Score:2)
Easy to imagine (Score:2)
What is 3 times better. Is 3 times better supposed to imply 3 times fewer crashes?
A grandma with poor eyesight can still drive around town without hitting anything, just slowly and cautiously.
A professional sports driver is many times better and can move the car a bit quicker because they are better able to handle changes. They also do not hit anything...
Do you now understand how a virtual "driver" might be three times (or more) better?
Show me the LIDAR and/or stereoscopic cameras (Score:2)
Show me where the LIDAR and/or stereoscopic cameras are.
I'll wait.
Can't have fully self-driving cars without at least one. You might be able to get away with Cadillac's high-definition maps but you didn't say that.
Re: (Score:2)
Show me where the LIDAR and/or stereoscopic cameras are.
There are 3 forward looking cameras in the windshield behind the rearview mirror.
There is no such thing as a "stereoscopic" camera. They are normal cameras and you combine the images to create an internal 3D representation.
Re: (Score:2)
Show me where the stereoscopic cameras are.
They are integrated into the main mirror.
Fleet and Sweep (Score:1)
The sooner they get a fleet of them on the road, the sooner we can run a sweep on how long it takes 4channers to have them driving in swastika formation.
It's called redundancy (Score:5, Informative)
If there's one thing I've learned from my dad, it's that no matter how good the software might be, sensors get faulty (especially in cars), and software, no matter how smart developers think they are, can never fully compensate for bad data fed by faulty sensors.
That's why everything is a Tesla is redundantly supported by other sensors - and the computers are redundant as well.
unless there's so much redundancy the system won't be affordable".
Tesla ALREADY HAS [fool.com] that level of redundancy. Plus of course, if too many sensors or compute is out they simply tell the driver self-dirivng is disabled and make them drive.
From the 2016 article:
"The new set of sensors going into Tesla vehicles in order to eventually make this possible include eight surround cameras for 360-degree visibility with as much as 250 meters of range, 12 updated ultrasonic sensors to add redundancies to the cameras, and forward-facing radar (which Tesla recently enhanced to serve as a primary control sensor alongside Tesla's vision system)."
Just as important as redundancy of sensors it has, is the fact that it also has multiple kinds of sensors which all combine data to understand the road ahead...
Re: It's called redundancy (Score:1)
So what you are saying is that it's redundancy all the way down?
Re: (Score:2)
Supposing there are layers upon layers of redundancy meant to stop a catastrophic accident, increasing the number of electronics increases the number of modes of failure for which there's no way to test. If Tesla can't even get hands free Bluetooth down [philly.com], how are we supposed to believe Musk's clai
Re: (Score:2)
OMG! We forgot that those electrons wear out after a few years of continuous use.
The purpose of redundancy is to provide backups if some electrons get tired and stop working.
You're right. It all depends on Bluetooth. No Bluetooth, no navigation. Damn vikings!
Re: (Score:2)
Re: (Score:2)
ICE cars have constant problems with sensors since they operate at high temperatures.
EVs don't have these high temperatures. Cameras only have to deal with ambient temperatures.
Re: (Score:2)
Of course, a Tesla is probably going to be junked after ten years or so when its battery really starts to fail and it's uneconomical to replace a 500kg battery.
Re: (Score:2)
Re: (Score:2)
Re: The big question... (Score:1)
They are cashing in on slashdot.
Who would have ever thought it could be managed worse than Dice ran it?