Samsung building autonomous chip for Tesla?

[electronblue]

The schedule for existing car FSD upgrades may not be as aggressive as some hope. It took a year+ for the sold-on-delivery P100D upgrades to arrive for late P90D cars and some promised retrofits we are still waiting for...

H1/2019 for HW3 hitting new production cars sounds plausible though.

 

[strangecosmos]

No idea whether Hardware 3 will mean big software improvements on day one, or whether it will take months. But here’s a key quote from the Q3 earnings call from Andrej Karpathy:

“Hi, everyone. My name is Andrej Karpathy. I'm the Director of AI here at Tesla. And my team trains all of the neural networks that analyze the images streaming in from all the cameras for Autopilot. For example, these neural networks identify cars, lane lines, traffic signs, and so on. The team is incredibly excited about the upcoming upgrade for the Autopilot computer which Pete [Bannon] briefly talked about.​

This upgrade allows us to not just run the current neural networks faster, but more importantly, it will allow us to deploy much larger, computationally more expensive networks to the fleet. The reason this is important is that, it is a common finding in the industry and that we see this as well, is that as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity.​

So in other words, we are currently at a place where we trained large neural networks that work very well, but we are not able to deploy them to the fleet due to computational constraints. So, all of this will change with the next iteration of the hardware. And it's a massive step improvement in the compute capability. And the team is incredibly excited to get these networks out there.”
​

My hunch is that within 6 months of the first 100,000 cars with Hardware 3 hitting the road, we’ll see at least one or two significant updates, whether that’s the option for unconfirmed lane changes for Navigate on Autopilot, highway on-ramp merging, stop sign warning/automatic stopping, red light warning/automatic stopping, automatic right turns at green lights and stop signs (probably with driver confirmation initially), or something else. This is just a gut feeling based on weak evidence.

I’m personally looking forward to July 2020, when Elon is 6 months late on his prediction to deliver full autonomy. That’s when things will really start to get interesting.

 

[CarlK]

HW3 is there only because of need for FSD. All new cars will get the same processor but EAP likely will not be getting capabilities beyond "freeway on ramp to off ramp" navigation. There probably will not be much change at HW3 release until FSD is activated. That's my interpretation.

 

[Bladerskb]

Let me remind everyone of Elon's prediction before re-visionary historians try to change it. Because July 2020 will make Elon up to 30 months late.

December 2015: "We're going to end up with complete autonomy, and I think we will have complete autonomy in approximately two years." -

Elon Musk Says Tesla Vehicles Will Drive Themselves in Two Years

January 2016: "In ~2 years, summon should work anywhere connected by land & not blocked by borders, eg you're in LA and the car is in NY" -

Elon Musk on Twitter

June 2016: ""I really consider autonomous driving a solved problem," he said. "I think we are probably less than two years away." -

Two years until self-driving cars are on the road – is Elon Musk right?

March 2017: "I think that is about two years" -

Transcript of "The future we're building -- and boring"

 

[Derek Kessler]

Let me remind everyone of Elon's prediction before re-visionary historians try to change it. Because July 2020 will make Elon up to 30 months late.

This is exactly why I say to never put any stock in Musk's projections.

 

[Vitold]

So far Musk has shown uncanny ability to predict trends and to pick right technology.

Going with their own chip is one of them. If in fact they partnered with Samsung that is even better. I always thought that they would go with AMD but Samsung could offer better/cheaper development for next gen chipset.

 

[CarlK]

So far Musk has shown uncanny ability to predict trends and to pick right technology.

Going with their own chip is one of them. If in fact they partnered with Samsung that is even better. I always thought that they would go with AMD but Samsung could offer better/cheaper development for next gen chipset.

That's indeed how his first principle of engineering design has worked. He correctly saw from the beginning brain instead of eyes is what we needed most for this to work. However I don't think Samsung has anything to do with design of the AI chip. At most its foundry is to fabricate the Tesla designed chip. AMD does not own a foundry anymore.

 

[ChrML]

It makes sense to use Samsung to produce the chips according to Tesla's design.

What would not make sense is building a productionline inhouse to produce chips only used by Tesla. This would be really expensive and require a very specialized people and equipment. Producing chips is not easy.

 

[CarlK]

Everyone except Intel uses those foundries. Even for Intel making chips only in house is one reason why it is starting to fall behind AMD and Nvidia.

 

[Vitold]

Everyone except Intel uses those foundries. Even for Intel making chips only in house is one reason why it is starting to fall behind AMD and Nvidia.

I do not thing anyone would expect Tesla to manufacture the silicon chip.

What Samsung or AMD could bring to the mix is integration with ARM core(s), hardware video processing etc

 

[boonedocks]

Are we talking 14nm or what? Also, the tone of Tesla and everyone assumes the design has been taped out and ready for mass production but I wonder why they aren't specifying anything about retrofit timeline vs. new production access to HW3? I have a bad feeling existing owners will get shafted in favor of new customers despite paying thousands years ago for this and the software for FSD.

I am sad to say I think croman is right. I have tweeted Elon twice with the same question with no response:

 

[croman]

I am sad to say I think croman is right. I have tweeted Elon twice with the same question with no response:

View attachment 363518

Sadly he doesn't care. Its all about pushing out new cars. We'll get the hardware, just whenever Tesla feels it doesn't impinge on their ability to make a profit. They might even tie it to an annual service like they did with uncorking. Whatever they decide, you can be sure it will be with an eye towards profit.

 

[mongo]

I am sad to say I think croman is right. I have tweeted Elon twice with the same question with no response:

View attachment 363518

Untill FSD is released, there is no reason to update HW2.x computers. (Unless it turns out EAP requires HW 3)

 

[zanary]

Untill FSD is released, there is no reason to update HW2.x computers. (Unless it turns out EAP requires HW 3)

The part that I think many are forgetting is that FSD is not a single feature, nor a single software update. It's a collection of features, and multiple software updates.

We know that the 1st feature of FSD will be coming in the 1st Half of 2019 and it will require the AP3 hardware (hence Elon's comment about the cars off GA will come with it, but I have to assume that FSD owners will also get it shortly thereafter). We don't know what that exact feature is, but we can extrapolate that it has to do with the neural networks because of the computational constraints of AP2/2.5.

[

So in other words, we are currently at a place where we trained large neural networks that work very well, but we are not able to deploy them to the fleet due to computational constraints. So, all of this will change with the next iteration of the hardware. And it's a massive step improvement in the compute capability. And the team is incredibly excited to get these networks out there.”
​

If I'm to guess what AP3 will give us...
(1) All of the cameras as part of the Dashcam solution. Right now the filename contains the word "front" so I assume we'll see the remaining cameras as well.
(2) MCU/2 will be required, but I guess we'll get a viewer for the Dashcam files as well
(3) Auto wipers that use the neural network (Elon said it's coming).
(4) Ability to read speed limit signs, understand street lights, etc... He said he's testing it, and I assume these use the advanced neural networks, so that might be a features of FSD.

I'm sure that many will not agree that FSD is a collection of feature, but if you look at EAP, it's not a single feature, but a collection of features that make up EAP. I have to assume that FSD is the same, and that EAP is done (refinements coming yes, but nothing new), and that everything going forward needs the AP3 computer and that's what Elon is already testing.

 

[Vitold]

Lol. How easy we forget trials and tribulations when 2.0 was first released.

As for Tesla it would be wasteful to retrofit old cars with new system while at the same time selling new cars with 2.0

 

[zanary]

Lol. How easy we forget trials and turbulations when 2.0 was first released.

As for Tesla it would be wasteful to retrofit old cars with new system while at the same time selling new cars with 2.0

I think the original owners of AP1 also had trials and tribulations as theirs was also delayed I think 6 or 8 months... History is History, and we're now in a spot where AP2/2.5 for EAP is pretty decent, and we are on to FSD for all the new cool stuff.

As for wasteful, I don't think they see it that way. They are obligated to provide current FSD owners with the hardware as per the T&C for the purchase of FSD. They'll be manufacturing enough for general assembly and at the same time to retrofit existing FSD fleet at a slow pace. I think they won't allow anyone to purchase FSD until the majority of the current fleet has been retrofitted... It also remains to be seen what the new FSD price will be.

Early Adopters got it at a serious discount, but everything else I think will need to pay up as it's worth more than double what early adopters paid for it.

Again, just speculation on my part of course.

 

[mongo]

The part that I think many are forgetting is that FSD is not a single feature, nor a single software update. It's a collection of features, and multiple software updates.

We know that the 1st feature of FSD will be coming in the 1st Half of 2019 and it will require the AP3 hardware (hence Elon's comment about the cars off GA will come with it, but I have to assume that FSD owners will also get it shortly thereafter). We don't know what that exact feature is, but we can extrapolate that it has to do with the neural networks because of the computational constraints of AP2/2.5.

[

If I'm to guess what AP3 will give us...
(1) All of the cameras as part of the Dashcam solution. Right now the filename contains the word "front" so I assume we'll see the remaining cameras as well.
(2) MCU/2 will be required, but I guess we'll get a viewer for the Dashcam files as well
(3) Auto wipers that use the neural network (Elon said it's coming).
(4) Ability to read speed limit signs, understand street lights, etc... He said he's testing it, and I assume these use the advanced neural networks, so that might be a features of FSD.

I'm sure that many will not agree that FSD is a collection of feature, but if you look at EAP, it's not a single feature, but a collection of features that make up EAP. I have to assume that FSD is the same, and that EAP is done (refinements coming yes, but nothing new), and that everything going forward needs the AP3 computer and that's what Elon is already testing.

I get what you are saying about sub-functions, but I'd say you can't have partial Fully Self Driving. Even level 3 is "driver present not needed" fuctionality. Given their development approach (single NN/ codebase), I don't know that they would even have release versions with subsets of functionality.

Speed limit sign reading was in AP1 so should be part of EAP.
On property summon will be EAP, summon via public roads is FSD.
EAP still needs improved stopped vehicle detection (unless that has resolved on a recent version)
Wipers use NN already.
EAP has speed control and auto steering (level 1 and 2 functionality), with the new NoA. FSD will need improved versions of all 3.
Given this tweet (bold added) I think Elon is referring to the current HW set.

Already testing traffic lights, stop signs & roundabouts in development software. Your Tesla will soon be able to go from your garage at home to parking at work with no driver input at all.

NN capability is a separate hardware set from multi camera dash cam. I'd guess the only holdup there is software.

 

[Electroman]

t’s in Tesla’s best interest to not become dependent on any single supplier.

OT: Yes and in a lot of ways they are dependent on Panasonic for both 18650 and 2170 cells. If for some reason Panasonic decides to part ways suddenly ala Mobile Eye, then Tesla is in deep feces trouble.

 

[Electroman]

Let me remind everyone of Elon's prediction

Thank you. I had almost forgotten all those predictions from Elon. I don't know what we will do without you. Keep up the good work.

 

[CarlK]

OT: Yes and in a lot of ways they are dependent on Panasonic for both 18650 and 2170 cells. If for some reason Panasonic decides to part ways suddenly ala Mobile Eye, then Tesla is in deep feces trouble.

That's one reason why Tesla always go for vertical integration. Don't think it will dump Panasonic as it dumped Mobilieye though. Yes desipit what Mobilieye wanted people to believe it was never in Tesla's long term plan. It knew that too. At most it accelerated the separation after learned Tesla's intention of eventually to do it alone. The Florida accident just provided a convenient excuse for not getting a too sever hit on the stock price. When your largest and most recognized customer wants to move faster companies would try to comply instead of dump it.

This was in the news long before the actual separation.
Elon Musk offered a “multimillion-dollar bonus” for Geohot to build a “Mobileye crushing” Autopilot system for Tesla