Tesla Is Building Its Own AI Chips For Self-Driving Cars (techcrunch.com) 157
Yesterday, during his quarterly earnings call, Tesla CEO Elon Musk revealed a new piece of hardware that the company is working on to perform all the calculations required to advance the self-driving capabilities of its vehicles. The specialized chip, known as "Hardware 3," will be "swapped into the Model S, X, and 3," reports TechCrunch. From the report: Tesla has thus far relied on Nvidia's Drive platform. So why switch now? By building things in-house, Tesla say it's able to focus on its own needs for the sake of efficiency. "We had the benefit [...] of knowing what our neural networks look like, and what they'll look like in the future," said Pete Bannon, director of the Hardware 3 project. Bannon also noted that the hardware upgrade should start rolling out next year. "The key," adds Elon "is to be able to run the neural network at a fundamental, bare metal level. You have to do these calculations in the circuit itself, not in some sort of emulation mode, which is how a GPU or CPU would operate. You want to do a massive amount of [calculations] with the memory right there." The final outcome, according to Elon, is pretty dramatic: He says that whereas Tesla's computer vision software running on Nvidia's hardware was handling about 200 frames per second, its specialized chip is able to crunch out 2,000 frames per second "with full redundancy and failover." Plus, as AI analyst James Wang points out, it gives Tesla more control over its own future.
Different headline than I expected (Score:4, Interesting)
From the previous thread about Tesla, I expected this headline to read "Tesla is now building their own arcade cabinets".
Re: (Score:2, Insightful)
It was either going to be "Elon's folly: why his chips wont work and he will become homeless trying to make them", or "Elon Musk set to disrupt the modern world as we know it with new Super-Computer on a chip stroke of genius".
What I can say is, the reason Tesla will be the biggest innovator and market leader in their field is simple. People are passionate about it. Good or bad, everyone has a strong opinion. Tesla these days reminds me of the "Pray" cover WiReD published about Apple before the second comin
Re: (Score:3)
Tesla has invented the fpga? Really? Why is,this exciting and ground breaking?
They do not own a fab or have the cpu designers that Intel, amd and invidea have.
This is just more Musk nonsense.
In 3-4 years, if tesla still exists, we will hear how the new ai chip is almost here. Just put down your non refundable deposit for the upgrade when you buy your car which will also be late and Musk will give you themupgrade 5 years later.... long after every other car maker has already perfected level 5 self driving vehicles and sells them at a million a week.
Andmtheir cars wont murder people either.
Yeah, right. FPGA is the technology they are espousing here. You are right on the button. I wish I was as insightful as you are. /sarcasm
They are investing in new applications of technology. They are doing their data collecting, determination of the best solution, and spending the capital required to further their lead. They don't have a FAB? Excellent! You only want a dedicated FAB when the quantities required are in the hundreds of millions. Why would they want a FAB? Apple designs the A(n) chips and uses
Re: (Score:2)
Unlike you, he doesn’t appear to be a vitriolic troll.
But he might have stock too, I don’t know.
Re: (Score:2)
I do not own any Tesla stock, nor are my comments paid for in any way. Don't come here playing identity politics with me. Who I am doesn't matter. What I am saying matters. Either my ideas are good or bad. If you want to refute them, refute them. But you do not know me, you know nothing about me. You are just a little retard who wont even sign up for an account here, little coward, little bitch.
Now, for my argument:
Elon is running two VERY VERY technical companies with insane technical achievement trajector
Re: (Score:2)
You are the biased one. Because the Merlin engine isn't that advanced - it is more or less 1960s technology but lighter thanks to the more modern materials becoming available since then. Landing the rockets is revolutionary, the engines aren't, though - not even close.
Re: Different headline than I expected (Score:4, Insightful)
Tesla has invented the fpga?
This is not an FPGA. It is a matrix math engine, like Google's TPU [wikipedia.org].
They do not own a fab
Other than Intel, nobody owns their own fabs anymore.
You just code up your chip in Verilog, debug in a simulator, and upload it to TSMC.
or have the cpu designers that Intel, amd and invidea have.
Neither did Google, but their TPU is a big success.
Re: (Score:1)
That'll be made by his Boring company.
Maximum Overdrive 2: Revenge of the Dissed (Score:5, Funny)
You know if you've ever said anything nasty about Elon.
Now, his vehicles know.
Be afraid. Be very, very afraid.
Re: (Score:2)
You know if you've ever said anything nasty about Elon.
Now, his vehicles know.
Be afraid. Be very, very afraid.
Wouldn't that kind of murder be more likely be perpetrated by the government or organized crime? if theres a difference between the two, that is
Re: (Score:2)
Of course there's a difference. Organized crime is - see the clue in the name - organized.
Seafood platter, etc.
Re: (Score:2)
I love the quaint belief that Americans have that corporations are more trustworthy than the government.
Re: (Score:2)
Re: (Score:2)
You can choose to live in a country or not. At least you can vote out a government.
Re: (Score:2)
Re: (Score:2)
What corporation can force you to work with it? Name one.
Private prisons :)
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
Guys that are serious about software, build their own hardware. - Alan Kay
Re: (Score:2)
This is actually a brilliant play. If they roll their own, it will be a differentiator his competitors wont be able to catch up with him on. They will do it themselves, be better at it through investment and expertise, having learned more by having billions of miles driven by their customers to teach their back-end system how to act. They will be market leading in driverless vehicles.
Re: (Score:2)
Except his chips will probably be differentiated by being more expensive and not as good as the chips made by people who have been doing it for years much like Tesla's production lines.
Re: (Score:2)
he's building (and selling) 5000 3's, 1000 S's and 1000 X's vehicles per week.at 20% margin(someone extrapolated a 30% margin on 3's see https://www.youtube.com/watch?... [youtube.com] ), for those 45000/100000/120000$ vehciles that would mean a total margin of 89 million$ per week, 4,6 billion per year margin. I believe Bernie would be envious about that.
Re: (Score:2)
Re: (Score:2)
Building proprietary silicon could be dangerous (Score:5, Interesting)
For better and worse, keeping things proprietary means it's by definition both closed source, and tested only to one's own environment. Although it produces fast yields, it doesn't have many eyes. Many eyes and many hours are needed to vet the integrity and edge cases (like cliff edges) before safety can be assured.
It's a risky, expensive, and proprietary endeavor. If everyone (systems builders) were using similar development, the testing age could be completed in a concurrent time, rather than a serial/iterative time. I'm betting against this turning out well.
Re:Building proprietary silicon could be dangerous (Score:5, Interesting)
Neural net calculations are pretty simple, just repeated many times over. Testing the silicon should be relatively simple compared to general purpose CPU or even GPU design.
Re: (Score:2)
First, you have to QA the silicon in all of the ambient environments, which means the worst of Alaska to the heat of Death Valley, all precip, all RF/EMI, etc.
Then using a closed OS (or a modified BSD-licensed RTOS), you have to ensure its reliability and recoverability from miscellaneous events. Then you have to train the sucker.
You can look to the failures of Google Maps for problems, including map-to-net errors, driver foolishness, and the craziness of random events, like rabbits and pedestrians, etc.
Her
Re: (Score:2)
First, you have to QA the silicon in all of the ambient environments, which means the worst of Alaska to the heat of Death Valley, all precip, all RF/EMI, etc.
Sure, but that's routine part of the design. Every modern device has several pieces of silicon that went through that process. There are tools and people that know how to do this.
Then using a closed OS (or a modified BSD-licensed RTOS), you have to ensure its reliability and recoverability from miscellaneous events. Then you have to train the sucker.
Which has nothing to do with proprietary nature of the silicon. Even if they had bought standard silicon, they'd still have this problem.
Re: (Score:2)
And all of this takes time, money, and both together equals margin, something that's pretty slim for Tesla.
Re: (Score:2)
Nothing of the like. I've watched wafer operations over the year hit the wall time and again. Fab shops aren't cheap. Renting a fab shop after you've re-invented the wheel isn't cheap, either. QA isn't cheap. Generations of this aren't cheap.
Re: (Score:2)
First, you have to QA the silicon in all of the ambient environments
I beg to differ, Elon's style seems to be to half-ass it and then tell his customers to use at their own risk.
Re: (Score:2)
>> I beg to differ, Elon's style seems to be to half-ass it and then tell his customers to use at their own risk.
Nope. They do the QA on the component level.
But they let the customers do some of the vehicle level testing.
Re:Building proprietary silicon could be dangerous (Score:5, Insightful)
Every NN is proprietary, and that is where the functionality to worry about is at. The performance on "edge cases" in driving is directly related to how much compute power you can throw at it. Tesla is multiplying its compute power. The edge cases will improve. Staying with the general purpose GPU instead of true NN hardware will guarantee continued unhandled edge cases.
This HW is undoubtedly also more energy efficient. That is the real key. They could stack on more boards, but these units are already consuming a significant amount of the vehicle's energy. The trick is to get more compute power with the same or less energy. NN specific HW is going to be a requirement to have that happen.
Everyone in the industry has known that GPUs will not be used past the first generation or so. They are development HW. Someone will eventually come up with a general purpose NPU that will win the market, but it hasn't happened yet - mostly because NN implementations haven't settled.
Re: (Score:3)
No, not yet.
Tesla PURPORTS to be able to add more NN intelligence using proprietary silicon. None of this is proven yet, just like many other Tesla missed goals. This is really rocket science, even with NN. Your rocket is an auto navigating known/unknown obstacles through its use cycle. It can't talk to other cars so as to update them with info about how squirrels leap out in front of you, then go back suddenly. The car will swerve anyway. Kill the squirrel is my opinion, personally, but these cars don't le
Re: (Score:2)
Re: (Score:2)
The cars themselves don't learn, but that would be very inefficient anyway. Tesla can review accident logs, modify their NN, and then send updates to all the cars.
FPGA technology makes no sense. FPGAs are slow and power hungry for NN applications. Besides, there's no reason to change the hardware. The same silicon can implement a wide variety of neural net topologies and weights.
Re: (Score:2)
FPGAs make great sense for execution. If you use a kernel that learns in a small environment and execute in an FPGA for routines, you get the limbic/pre-frontal cortex analog. FPGAs can run circles around GPUs as well, given the scope of execution needed.
As far as reviewing accident logs, that's much tougher. Ingesting the logs to discern actions to take/conditions to learn becomes not so much arbitrary, but a difficult regimen itself. And it means that it's forensic, rather than preventative. Knowing what
Re: (Score:2)
FPGAs can run circles around GPUs as well, given the scope of execution needed
You have no idea what you're talking about. NN evaluation requires multiplications, big caches and fast external memory access. FPGAs are lousy for all of those things. Big FPGAs are crazy expensive too. And their key property, the ability to upgrade them in the field, is totally useless for this application.
Re: (Score:2)
NNs learn. FPGAs are easily sent through (new) subroutines. Wanna do algorithms? Field programmable ones, ones that don't have to wait for storage? And they're not that expensive. As a microcontroller in this app, they're great for sensor monitoring.
Whatever the actual design that arrives, it's proprietary, and isn't going to be open to scrutiny, and will be expensive. I still believe it's a bad move.
Re: (Score:2)
No Neural Nets are “trained” in the lab. Then you choose the best one for your deployment, and make it as efficient as possible. The th8ng you send out into the field is quite static. It no longer “learns”.
Re: (Score:2)
Re: (Score:2)
They're not fully autonomous systems, and the data that they gather is not from fully autonomous systems. Those I know driving Teslas like to actually drive them.
Dissimilar software runs now on Tesla cars as well. Amalgamating that data is no easy task. And we agree that this amalgamation is important.
This is why I'm personally advocating for a unified approach to the problem, believing that it will eventually be a money pit for Tesla.
Re: (Score:2)
Problem is they already sold hundreds of thousands of vehicles with the full self driving option, so they will all need to be retrofitted with this chip if that's what self driving requires. And I'd guess more sensors too, because the cameras they have are probably inadequate and don't have any self-cleaning functionality that will be essential.
It's no wonder they jacked up the price of the self driving option.
The whole thing is another lawsuit waiting to happen.
Re: (Score:2)
Maybe. I suspect they just customized the pipeline here to keep things from having to go to external memory and come back in. Human vision processing works in that way. There are several stages of specialized hardware producing results and sending them off to the next stage which operates in parallel. The early stages don't need NNs at all. The NN usage increases as you get to later stages. Using the same HW for every stage means that there is a lot of wasted HW. It also often means that multiple stages tha
Re: (Score:2)
By going baremetal, they're reducing complexity and eliminating most of the potential for what you're talking about. Also, I'm guessing this might be a realtime system, in which case tolerances will be even tighter and bugs would be more obvious and thus more likely to get identified.
Re: (Score:2)
it's by definition both closed source, and tested only to one's own environment
So basically the entire self driving car industry with the concept of custom silicon being completely irrelevant?
If anyone needs me... (Score:5, Funny)
Re: (Score:2)
Good idea. Especially if they are Uber cars.
Surely a bad decision (Score:4, Insightful)
"We had the benefit [...] of knowing what our neural networks look like, and what they'll look like in the future,"
Really? If they take their neural network development seriously I don't think they know what their networks will look like in ten years. It's a research area in the middle of a transformation. Using architectures molded into hardware is probably just costly and will act as an antagonist to innovation. I don't think having 2000 vs 200 frames per second right now outweighs that downside.
Re: (Score:1)
Re: (Score:2)
I think you will find that should be "She":
https://teslamotorsclub.com/tm... [teslamotorsclub.com]
Re: (Score:2)
Re: (Score:3)
There's no guarantee that off-the-shelf hardware from NVidia is a better match for networks 10 years from now.
Re: (Score:2)
... and if Nvidia is better in 10 years, Tesla will have the option of buying Nvidia hardware at that time. Developing in-house hardware no doesn't mean they can only use in-house hardware for the rest of time.
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
To what end? (Score:4)
A car travelling at 90 mph is moving about 4 cm/millisecond. So going from 200 fps to 2000 fps is going from 20 cm to 2 cm per cycle. What's the use of recognizing a car every two centimeters? For a jogger at 9 mph it's down from 2 cm to 2 mm. It's neat and all but I don't see how that necessary to react in the time frames a car needs to react. Even if it takes 3-4 frames for the car to get a motion vector 0.2 seconds is still way quicker than a human and 0.02 seconds doesn't bring that much.
Re: (Score:2)
Re:To what end? (Score:5, Interesting)
Maybe after a point, but up until that point, the main risk is reacting too slowly. Ask anybody with an AP2 Tesla how well it handled curves prior to earlier this year. Of they don't use the word "lag", they don't know software, and if their eyes don't bug out in abject terror, they don't know how to drive.
Basically, it had (and still has, to a lesser extent) trouble with lane keeping, because its reactions lagged behind reality, and it started turning way too late, resulting in uncomfortable turns, getting dangerously close to barriers and center lines, etc. This is better in current versions, but I still get scared enough to take manual control a couple of times per day.
So right now, performance is still their main problem. This is a very welcome announcement.
Manual handling (Score:2)
Very practical question regarding Tesla handling (never drove a Level 2+ car before, only my parents' car with FCAS and LDWS, and various similar system on all the various fleets of car-sharing).
Normally, a Telsa should handle both steering and accelerating/braking.
I the driver where to grab the wheel to adjust steering - BUT keeps the feet hovering above the pedals without pushing them (Yet) - would this disengage only the autopilot steering ? While keeping engage the usual distance keeping / emergency aut
Re: (Score:2)
If you press the gas pedal too much, you'll disable automatic braking, and nothing else. If you touch the brake, it disengages everything. (This is, IMO, a bug; it is too easy to accidentally tap it on a curve and have the wheels s
Re: (Score:2)
Most drivers never tried to use AutoSteer in places where it mattered. Those of us who drive CA-17 and other similar roads regularly were the only ones swearing about it, but we were very much swearing.
And I made fairly public comments here on Slashdot and Tesla discussion boards about it, noting that it was unlikely that the software being used for AutoSteer at the time was actually the base code for their self-driving feature, but rather temporary code that they were patching and tweaking minimally, in m
Re:To what end? (Score:5, Insightful)
Re: (Score:1)
Think of a video game. If you dial your resolution and detail back to 1990s levels, most modern GPUs could perform thousands of FPS on the top modern games. They don't because we've used the rate gain to produce higher fidelity and more detail.
Same here. It may run their current NN at 2K frames per second. But they will change that network to utilize the extra time to get better results than they currently have at 200 FPS.
Think of the NN as a Chess engine. It can consider more with more compute power or it
Re: (Score:1)
Re: (Score:2)
No company wants to be the coachmaker, placing their design over another brands "computer".
Re:To what end? (Score:4, Interesting)
That's the answer they want people to hear. The real answer is no longer having to pay nvidia.
Re: (Score:2)
0.2 seconds is nearly two car lengths at 90mph
Re: (Score:2)
Re: (Score:2)
A car travelling at 90 mph is moving about 4 cm/millisecond. So going from 200 fps to 2000 fps is going from 20 cm to 2 cm per cycle. What's the use of recognizing a car every two centimeters? For a jogger at 9 mph it's down from 2 cm to 2 mm. It's neat and all but I don't see how that necessary to react in the time frames a car needs to react. Even if it takes 3-4 frames for the car to get a motion vector 0.2 seconds is still way quicker than a human and 0.02 seconds doesn't bring that much.
2cm is about 3/4 inch. Could the vehicle recalculate and take a trend line view of a problem. That would lead to sufficient leadtime to prevent a fatality. If the car could know what to do by one foot of travel (at 90mph), wow.
Existing owner (Score:1, Funny)
Seriously : Yes, (hardware) upgradeable (Score:2)
There's two components on an autonomous system.
The sensors and the computer.
Tesla has publicly stated that they've on purpose designed the computer to be modular, and the current cars since recently (forgot the exact date, it's google-able, I think it's since the introduction of the triple front cam) are designed in such way that you could swap the computer with a newer one in the future.
So in theory yes, if they keep their word, you should be able to install the newer computer with the better NN when it's
And then something changes... (Score:2)
What happens if (or, when) Tesla realizes they need to make a significant change to their code?
Automated driving and AI are both hot research areas. I wouldn't take a bet that there won't be big changes in the near future.
This smells like an unholy combination of two things: a development team getting burnt by premature optimization, with just a hint of "painting oneself into a corner".
Between this and the omission of lidar, I'm not enthusiastic about Tesla's self-driving capability. My pessimism applies ac
Re: (Score:2)
Re: (Score:2)
I wouldn't take a bet that there won't be big changes in the near future.
Most likely, the silicon can handle future changes just as well as current graphics cards, if not better. Also, I would expect Tesla to keep evolving their hardware.
Re: (Score:2)
Between stereo and inter frame processing using the known car speed/direction you can reconstruct a 3D representation without LIDAR. LIDAR has problems too. The mutual interference which LIDAR suffers from requires some very expensive tricks to prevent ... especially expensive for non scanning LIDAR, which is the best LIDAR because fuck mechanical 2D scanners. LIDAR is a dead end. high resolution radar would be nice though for heavy rain and fog.
As for designing their own neural network ICs, better than be
Sensors : Getting there. Eventually. (Score:2)
This smells like an unholy combination of two things: a development team getting burnt by premature optimization, with just a hint of "painting oneself into a corner".
Well, on the other hand, It's just number-crunching hardware, used to run their NN.
Maybe they'll come up with a good computer, maybe in the future they'll realise that silicon by Nvidia or AMD ends up being the best to run their nets.
It's more important to them (they'll be divesting money into that R&D) than to users (it's just number-crunching silicon to run a NN on it. You're supposed to be able to swap the computer for a better one in the future, according to Tesla).
Between this and the omission of lidar, I'm not enthusiastic about Tesla's self-driving capability.
I totally agree with this. At a ti
Free upgrades, right? (Score:2)
As they have been advertising for a long time that what you buy now contains all the hardware required for fully autonomous driving with a free software upgrade in the future.
ASIC (Score:2)
Comparing to Apple? (Score:2)
Apple has to fit its chip into a 4 oz container slightly larger than a credit card. You've got an entire f'in car, put 100 chips in it genius.
Potential upside of neural nets (Score:1)
Ambitious (Score:1)
Why does he think he can start up a new business like that and eclipse Nvidia that has decades of experience? He's been the dog that caught the car. The dog can't keep doing that.
Could be the first (or second... or so on) nail in this coffin.