Did Elon admit on the investors call that card are going to need a new processor?

[chispas]

I haven't read anyone talking about it but it seemed huge to me. Elon was soft about it (might, maybe, etc.) but coming from him I think they are going to need it but they aren’t ready yet to put that in the cars.

You can listen Elon talking about it in the Tesla Q1 Webcast min 39. Transcription of what Elon said (I tried to be as close as what he said as possible, sorry if I missed something):

"I believe that the current production, the vehicles that are currently producing, are capable of full autonomy with the only thing that really be ... that might be needed ... maybe is probably needed is a computer upgrade to have more processing power for the vision neural net. But that is a plugin replacement, something that can be done quite easily"

 

[drewbaumann]

It is possible. If it does need an upgrade I believe he is correct about it being a swappable part.

 

[Discoducky]

He's said this before, but Drew is correct, fully replaceable and will be done for free.

 

[chispas]

The fact that is easily replaceable and that I assume is easy to do is good news. My reading is that the hardware that they want to use - maybe their own chip? - isn't ready yet, otherwise, why are they not putting it in already and save the labor later? Option b, maybe they just don't know for sure what is going to be needed for sure.

 

[drewbaumann]

The fact that is easily replaceable and that I assume is easy to do is good news. My reading is that the hardware that they want to use - maybe their own chip? - isn't ready yet, otherwise, why are they not putting it in already and save the labor later? Option b, maybe they just don't know for sure what is going to be needed for sure.

I would assume that the hardware (cameras and sensors) are autopilot ready and that the chips currently in the cars work well for gathering data and handling EAP. However, with AI technology improving rapidly it would be cheaper waiting until the software is ready to upgrade the cars. This gives them room to use their own chips, the latest and greatest from Nvidea, or some chip that is sufficient, but is new and expensive now, but they could get a deal on if they buy it after newer tech is available.

 

[croman]

Yea, this is old news. Elon keeps saying that. We knew that when AP2 was announced. They cheaped out and got the one chip version instead of what Nvidia said was the autochaffeur 2 chip board.

Presumably Keller's departure means their AI chip is taped and masked and ready to be fabbed. That will take 6 months (or so). If the chip design is flawed, it will be 12-18 months before the board can be respun and fabbed. Lets hope Keller nailed it and AK keeps killing it with his vision stack.

 

[chispas]

Yea, this is old news. Elon keeps saying that. We knew that when AP2 was announced. They cheaped out and got the one chip version instead of what Nvidia said was the autochaffeur 2 chip board.

Interesting, I've always heard that Tesla said that if the hardware wasn't sufficient they would upgrade it but never a clear "maybe is probably needed a computer upgrade" but I'm also kind of new (Dec 2017) so that is that

 

[croman]

Interesting, I've always heard that Tesla said that if the hardware wasn't sufficient they would upgrade it but never a clear "maybe is probably needed a computer upgrade" but I'm also kind of new (Dec 2017) so that is that

Elon Musk: Your Tesla May Need Computer Upgrade for Full Autonomy | Inverse

Musk Admits Tesla's Automated Drive System May Need More Computing Horsepower

Elon is admitting the power isn't enough but its very advantageous to wait as long as possible because EAP doesn't need more horsepower, just FSD and their new chip should be way cheaper than anyone else's chips and should be optimized for AI and not just a superpowerful GPU.

Of note, even in Tesla's official statement confirming the upgrade to the CPU, the upgrade will not happen for new other kinds of hardware (cameras, radar, etc.). Further, only people that pre-purchased FSD will get the upgrade. Its an option contract and it is very likely Tesla will raise FSD prices once the software is close at hand because the utility of FSD is far higher than EAP and therefore FSD can be priced higher and customers will likely still pay. Also note that Model 3 pricing for the software is the same as other models despite the percentage of the total cost being much higher. Tesla's bean counters clearly look at FSD as a discounted risk option currently and they will revise that later, particularly if there are added hardware costs to internalize for prior buyers (likely the $1k added cost option will be much higher for those that did not exercise the option before it is changed/rescinded).

 

[lux_cars]

Elon Musk: Your Tesla May Need Computer Upgrade for Full Autonomy | Inverse

Musk Admits Tesla's Automated Drive System May Need More Computing Horsepower

Elon is admitting the power isn't enough but its very advantageous to wait as long as possible because EAP doesn't need more horsepower, just FSD and their new chip should be way cheaper than anyone else's chips and should be optimized for AI and not just a superpowerful GPU.

Of note, even in Tesla's official statement confirming the upgrade to the CPU, the upgrade will not happen for new other kinds of hardware (cameras, radar, etc.). Further, only people that pre-purchased FSD will get the upgrade. Its an option contract and it is very likely Tesla will raise FSD prices once the software is close at hand because the utility of FSD is far higher than EAP and therefore FSD can be priced higher and customers will likely still pay. Also note that Model 3 pricing for the software is the same as other models despite the percentage of the total cost being much higher. Tesla's bean counters clearly look at FSD as a discounted risk option currently and they will revise that later, particularly if there are added hardware costs to internalize for prior buyers (likely the $1k added cost option will be much higher for those that did not exercise the option before it is changed/rescinded).

Also depends on what the completion is charging for FSD when it’s available.

 

[Canuck]

(likely the $1k added cost option will be much higher for those that did not exercise the option before it is changed/rescinded).

Once I see it working I won't mind paying more for it. Paying for it now looks like a class action hassle but I hope I'm wrong about that. In any event, I left the FSD box unchecked when I ordered my 3. I really need to see it to believe it and until then my money is staying in my wallet. But I'll be more than happy to pay more later -- I just think it's going to be a really long time later.

 

[Mo City]

Once I see it working I won't mind paying more for it...I really need to see it to believe it and until then my money is staying in my wallet. But I'll be more than happy to pay more later -- I just think it's going to be a really long time later.

I think a huge percentage of Model 3 reservation holders share your sentiment. If/when FSD becomes a reality (or is clearly imminent), Tesla will increase the price significantly and a large number of people will happily pay for it.

 

[jimmy_d]

It's an unsolved software problem at this point, so the amount of hardware you need to do it adequately isn't yet known by Tesla, or by anyone else. Not only is it not know in practice but there's also no theoretical foundation for arguing that any particular amount of hardware performance is required. NN development is thoroughly empirical at this stage. The existing GP106+Tegra X1 hardware might be perfectly adequate, or it might not.

How it unfolds will be interesting to see. The earliest software package is likely to take more hardware resources and perform less well than later versions. As the software matures it will take less hardware so the longer it takes to bring out the software the more likely it is to work within any particular hardware performance envelope. And the regulatory question is really hard to predict - just how good does an FSD system need to be before it'll be allowed on the roads? And how do you satisfy notoriously conservative regulators that your system is going to be acceptable to a public which is already primed to be skeptical by hyperbolic media accounts? If regulators delay release of the system by 5 years then the likelihood that it'll work on existing hardware is a lot higher than if regulators allow the earlier versions to go out to the public.

 

[McRat]

AFAIK the Level 5 nVidia autonomy board is the Pegasus, and it's not been released yet. It's a 16gb dual 'Volta' card with 320 trillion operations a second.

How much will it cost? That will be interesting. Price a nVidia V100 single "Volta" Tesla compute card. Yes, nVidia was selling Tesla cards before Tesla was selling Tesla cars.

Could be affordable. Might not be. No price announced yet.

 

[drewbaumann]

All this thread does is make me wonder if it's best to change my order to include the autonomous package for fear it will be drastically more expensive when released.

 

[S4WRXTTCS]

All this thread does is make me wonder if it's best to change my order to include the autonomous package for fear it will be drastically more expensive when released.

I'm not aware of a single self-driving car expert who believes that the current HW is capable of FSD. They're not even considering the vision computer because they assume that will get upgraded.

No one knows what regulatory requirements are going to require in terms of hardware.

So I think it's best to hold off until there is an actual path towards FSD.

 

[JohnnyEnglish]

It's an unsolved software problem at this point, so the amount of hardware you need to do it adequately isn't yet known by Tesla, or by anyone else. Not only is it not know in practice but there's also no theoretical foundation for arguing that any particular amount of hardware performance is required. NN development is thoroughly empirical at this stage. The existing GP106+Tegra X1 hardware might be perfectly adequate, or it might not.

How it unfolds will be interesting to see. The earliest software package is likely to take more hardware resources and perform less well than later versions. As the software matures it will take less hardware so the longer it takes to bring out the software the more likely it is to work within any particular hardware performance envelope. And the regulatory question is really hard to predict - just how good does an FSD system need to be before it'll be allowed on the roads? And how do you satisfy notoriously conservative regulators that your system is going to be acceptable to a public which is already primed to be skeptical by hyperbolic media accounts? If regulators delay release of the system by 5 years then the likelihood that it'll work on existing hardware is a lot higher than if regulators allow the earlier versions to go out to the public.

The April 17th tweet by Andrej Karpathy talking about how he reduced execution time for a block of code from 13 seconds to 0.02 seconds is a good example of the sort of optimisation that is possible.

 

[run-the-joules]

Also depends on what the completion is charging for FSD when it’s available.

Assuming you meant competition, but it feels like most companies chasing full autonomy are looking to provide vehicles-as-a-service, and it feels like only Tesla (and maybe Audi) are trying to satisfy those of us who prefer ownership and not-sharing.

 

[banned-66611]

He is just preparing Investors for the bad news that major sensor upgrades will be required. There is no way the current package is going to deliver what they have been selling people.

 

[WarpedOne]

Thank god for the ignore feature!
One of the most useful and nerve-preserving feature of this forum.
So far, the universe is winning.

 

[Tezlanian]

The requirements are unknowable generally since no one knows what the computing requirements will be until it gets closer. That said, it's highly unlikely the current "super computer" would ever have been capable in anyone's mind, realistically. The final version will no doubt be highly concentrated inference ASICs.