Forums

Tesla will start to use it's own AP chips next year.

Tesla will start to use it's own AP chips next year.

This will set it apart from the rest forever. I thought this is the even bigger news than 6,000 a week M3 production we hear today.

jjgunn | 1 agosto 2018

Tesla manufacturing hardware?!?!

Going to be bigger than Amazon

KP in NPT | 2 agosto 2018

I actually can't wait for more on this. This was probably the biggest story on the call and the press hadn't grabbed it as of last night.

I remember Elon saying a while back they were working on their own chip. Pretty sure it was sort of poo-pooed at the time. if they've done it and it can do all they say, this is going to put them way ahead in the autonomous race. A chip like this could be an entirely new company. Stock will go through the roof.

ff9700 | 2 agosto 2018

https://techcrunch.com/2018/08/01/tesla-is-building-its-own-ai-chips-for...

And they mention in this article that S/3/X computers will be swappable for the new "computers". Wonder if it would be a free upgrade...

matt80206 | 2 agosto 2018

Fingers crossed probably just for those who paid for full self driving though

mntlvr23 | 2 agosto 2018

.... and so it continues ..... Tesla improving upon cars that have already been sold and put into service - and shunning the model of "Planned Obsolescence"

RedPillSucks | 2 agosto 2018

@ff9700 I'd imagine it should be a free upgrade if you've paid for EAP or FSD.

KP in NPT | 2 agosto 2018

They said it's a free upgrade.

JAD | 2 agosto 2018

I have always thought the upgrade was free IF you chose FSD option. That was one of the main reasons to buy it in advance. It has always been stated that any needed upgrades for FSD would be included in the $3k FSD purchase. They have never said that for the EAP as the current hardware is likely able to handle EAP, just not FSD.

A 10x improvement with excess capabilities to what the software team needs is huge.

Alex_SD | 2 agosto 2018

One huge advantage Tesla has over the competition is the availability of a large amount of cars sending large amounts of data to data centers. This is the unique and critical to AP design and improvements.
However, they need to figure out how to handle the communication part. The low speed LTE available today is definitely not enough.
Besides the AP chips, they also need the new coming 5G modems.
Definitely looking forward to the Model Y with 5G! I don’t believe that they can retrofit the exiting cars with all these new technologies... especially 5G...

Mike83 | 2 agosto 2018

Looks like they let the cat out of the bag. I agree this is very important but don't mind if it the news misses it. Not sure how it will affect NVDA although I had sold it months ago and got more Tesla due to the Tesla vision.
A most impressive call.

Garyeop | 2 agosto 2018

I would think Tesla would save money developing against one chip. This would mean not only a free upgrade, but a mandatory one. It also makes them less valuable as a stolen car. Win win win.

Revelate | 2 agosto 2018

Chip design in house is not a panacea.

I don’t read that much into this long term look at how other semiconductor manufacturing has gone over the last several decades?

Still really early yet, and the regulatory hurdles for autonomous driving is non-trivial.

I think it is great that Tesla is pushing the envelope with this, and it may turn out to be huge, but I suspect the reason it hasn’t really been made much of in the media is simply this is a long way out; cool R&D effort but isn’t immediately applicable and there is a lot of effort from other companies with massive pockets who build their own computer hardware today doing similar (vis a vis Google/Waygo) so it isn’t an automatic world beater... even if it is now making me rethink FSD option if the August features are awesome.

Sadly, we still don’t know how changing options really affects timing and features (free LTE or free Supercharging or whatever) based on last config date vs initial $2500 for all that it has been repeatedly stated it doesn’t matter.

jjgunn | 2 agosto 2018

Mike83 | August 2, 2018
Looks like they let the cat out of the bag. I agree this is very important but don't mind if it the news misses it. Not sure how it will affect NVDA
----------
I got off my NVDA too - Just couldn't resist the profit

It's funny (SAD) how the media picks & chooses what they want to hear.

carlk | 2 agosto 2018

Elon has said it will be a free upgrade if the new processor is needed for FSD to work. I imagine it will only be for those who have purchased FSD option. Others may have to pay extra of pay for future upgrades. No matter what it's always a good thing to have the option available to you. Imagine what if you can swap out your iPhone board when a new version of processor comes out. This is the last piece of puzzle to keep your Tesla forever young.

Agree with Mike it does not matter if it's in the news or not. Some people may say Tesla is all hype but we here know it's all about results. When the result comes out in a year or two people will not be able to help but to take notice. As for NVDA it should be alright. No one else could do what Tesla is doing to design its own chips. I don't think Tesla will be interested in supplying chips to other auto companies to compete with NVDA.

ReD eXiLe ms us | 2 agosto 2018

I remember a rumor from 3 or 4 years ago that Tesla was working closely with AMD on some kind of AI chip. That may have been the behind-the-scenes beginning of their rift with MobilEye. It will be interesting to see how Autopilot continues to evolve over three platforms: MobilEye, Nvidia, and TeslaVision.

mntlvr23: Yes. I, too, prefer 'Planned Awesomeness'.

JAD: I suspect Tesla would like to make the upgrade process as smooth as possible. Since Autopilot works in 'Shadow Mode' to send data back to 'The Mothership' anyway, it may well be worth it to Tesla to retrofit cars that don't have activated EAP or FSD eventually, as a regular part of Service, once it becomes the new standard. Every data point helps with machine learning. Right?

carlk: Indeed. Tesla will surely get Full Self Driving working to an extent beyond what traditional automobile manufacturers manage using the Nvidia system first, before switching to their own chips. That will lead more and more tradit iona automobile manufacturers to choose Nvidia, once they realize that MobilEye is a dead end. So Nvidia will be just fine, supplying all the automobile manufacturers that Follow Tesla into a fully electric future. No need for Tesla to compete as a Supplier with Nvidia at all.

Of course, there may come a time, where if Tesla's cars are truly Autonomous and other vehicles have not progressed, stuck for five, seven, ten years or more... Tesla might decide to open source their system for the sake of safety in use in vehicles without ICE or HFCEVS drivetrains, at least. Y'know, the way Volvo did for the original design of shoulder belts.

Volvo Gave Away Their Most Important Invention To Save Lives | jalopnik.com
https://jalopnik.com/volvo-gave-away-their-most-important-invention-to-s...

Effopec | 2 agosto 2018

I wonder if this is the reason for the increase bump from $1k to $2k to buy FSD in the future?

carlk | 2 agosto 2018

If Tesla ever supply its system to others it sure will make it available only to BEV's. One reason that Elon began the self driving endeavor I believe is he wants to have every single EV Tesla produces to replace many ICE vehicles on the road. No idea what other automakers' intentions are other than it would not look good if they don't do this too. Having a lot of self driving ICE cars running 24 hrs a day on the road is not something we'd like to see.

Coastal Cruiser. | 2 agosto 2018

Yes, as you think about the idea of Tesla selling FSd solutions to other automakers it does make sense. They are playing into their [Silicon Valley] strength.

Did you hear Elon state that the current manufacturing issues are mostly software related? I wish he had broken that down a bit. But it appears he sees better code as the key to increasing existing line volume (as opposed to the letter statement that they are going that route as opposed to spending the cap on building new lines).

Coastal Cruiser. | 2 agosto 2018

ff9700 "https://techcrunch.com/2018/08/01/tesla-is-building-its-own-ai-chips-for..."

That was a very helpful link, as I missed this critical quote on the call:

“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.”

For non-techies, it may be helpful to point out that what Tesla appears to be doing here is a huge-o FIRST PRINCIPALS THINKING approach. The "GPU" or Graphics Processing Unit" is a specialized microprocessor designed to run many calculations at the same instant (aka Parallel Processing). The GPU was designed for gaming but is a perfect fit for self-driving technology (your brain is a GPU).

But the punchline is that Tesla does not appear to be building a better GPU. From what Elon said they are BYPASSING the GPU (which has to do some code translation, which takes time) by creating code that runs at the "bare metal level", which I assume means at the binary level, which is the native language of the machine.

And you can see why NVIDIA is out of the picture, as the GPU is its bread and butter.

Revelate | 2 agosto 2018

Actually a little simpler than that: basically the code will run natively without emulation.

Anytime you have something which isn't binary compatible (emulation), or has to be interpreted, it is a performance penalty and it doesn't matter what the application is. Admittedly, sometimes you can get good enough performance which is why Java took over the enterprise business logic space at one point even though it was absolutely slower than a native C++ or C and of course Assembly (which would be asinine to try to write in for that but I digress).

Essentially, they are building a custom chip specialized for this application, so a whole bunch of the generalized support that you'd find in a CPU or even GPU and just gets thrown away. Less transistors = less heat generated = faster clock cycles without frying the silicon, and being able to run the application natively on the hardware is another win too.

Nothing but goodness, but talking what a hundred thousand of these processors in the next year? That's not a threat to NVidia yet and they're not going to be idle either. Somewhat niche use case for now, though longer term there isn't that much difference between autonomous driving and more sophisticated applications (like a robot that could play dodge ball let's say) just a different application... which is where I think that one analyst was heading on possible military applications of the technology.

Just hard to say with any new tech where it goes long term, but definitely cool for me as an admitted geek.

carlk | 2 agosto 2018

Looks this is actually a neural network AI chip not just a processor. I'm not sure if anyone one else has it but I'm sure no auto company has it or will be able to make one in the foreseeable future if ever. We first saw signs of this almost three years ago but Tesla probably has been working on this long before that. It just has too much of lead for others to catch up now.

disapr | 3 agosto 2018

Generally what they are building is an ASIC (application specific integrated circuit). They have been developing this for a while so I think NVIDIA was just a stopgap. You can never have enough compute power, but this large of an improvement should enable FSD unless there is a fundamental flaw in the decision to use computer vision. I hope that the cameras are equally easy to swap since they might find that they have the headroom and necessity for greater resolution.

carlk | 3 agosto 2018

@disapr Sensor suit in HW2 is already better, probably way better, than human eyes. While everyone is working on sensors and follow Google with expensive LIDAR Elon knew from the beginning it's the brain that matters. It's true for human and true for the machine too. You don't need to improve drivers' eyesight when training them for challenging tasks like driving a race car. First principle engineering trumps again.

disapr | 3 agosto 2018

@carlk AP2 cars made in 2016/2017 will be using pretty old cameras on the timetable of technology by the time that the AP3 computer fully deployed and being retrofitted. I just think it's not unreasonable that before the not-yet-released AP3 computer is EOL, there could be a limitation in capability due to the 2016 spec vision sensor. Time will tell but my money says they are either implementing a new image sensor with AP3 or they will before it's EOL.

mos6507 | 3 agosto 2018

[I'm not sure if anyone one else has it]

They do. Google has Tensor chips.

My feeling is that if they are going to have the extra speed they should definitely upgrade the backup camera and incorporate the video feed into autopilot. The car really should have full 360' situational awareness including behind itself. Then it would know when it can jam the brakes on without causing a rear-ender with a tailgater.

carlk | 3 agosto 2018

@disapr Since the current sensor suit already has better capability than human eyes the bottleneck must be the "brain", both hardware and software part of it, at this moment. You can also look at it this way. If the neural net can reach the capability of human brain the car should be able to drive as good as the best driver by getting inputs from just two human eyes. From their actions apparently the Tesla team believe that too.

jjgunn | 3 agosto 2018

mos6507 | August 3, 2018
[I'm not sure if anyone one else has it]

They do. Google has Tensor chips.
-----------
That's nice & all but is Google turning into a car company now too? I know they're working on Waymo but they ain't making cars now.

Tesla is the only car company that has anything like this. That may change int he future but right now, they're the leader in the clubhouse by a light year