Tesla autopilot HW3

[thewishmaster]

Pete Bannon said on the Q3 earnings call that Hardware 3 should be "ready to go" by the end of Q1, i.e. the end of March. In October, Elon said "~6 months" on Twitter, i.e. ~April.

It would be cool if some or all Model 3s are already shipping with Hardware 3, but isn't the more likely explanation that Tesla simply changed its documentation ahead of time, in preparation for the launch in a few months?



Is this a typo or is info about Hardware 4 really already coming out?

They don't have 2.5 as a version in that list, so I'd guess APH3 = "HW2.5" and APH4 = "HW3"

 

[MorrisonHiker]

They don't have 2.5 as a version in that list, so I'd guess APH3 = "HW2.5" and APH4 = "HW3"

That's the assumption. We do know APH3 is AP 2.5 so one would assume APH4 is AP3.

 

[J1mbo]

One thing that MIGHT indicate HW3 on European cars is that NONE of the test/press cars that currently drive around in Europe have Autopilot activated. There is just standard cruise control. All test/press cars before have ALWAYS had Autopilot activated.

Reason might be that they have HW3 and the software is not ready for prime time yet and hence deactivated. In a few weeks when customer deliveries will start, they have to have it ready.

Just a theory though, but hope it is true!

I thought the EU press cars have all been US-spec cars? Certainly the showroom cars are all US-spec.

 

[HolyGrail]

Them US-spec cars wouldn't be deemed safe to drive in good ol Europe.

 

[PaulJohn]

I wonder if they wanted to optimize the news cycles or something. Maybe they do have something almost ready to announce and they want it to make all the headlines and act as another demand lever....

Bingo.

 

[evster]

I thought the EU press cars have all been US-spec cars? Certainly the showroom cars are all US-spec.

The showroom cars that arrived in Europe early december were all US-spec cars and had the US charge-port. Those cars were just to try out in the showroom and not legal for street driving.

Last week there were however an early batch of REAL european Model 3 arriving. Those cars all have VIN in the range of 192XXX with EU-spec (CCS-charge port, they don't have the orange small light in the front DRL (white in EU-version) etc.) Those cars are now just recently handed out to the press and the first EU reservation holders are test driving and all have AP disabled.

Linked below is a Dutch review that popped up yesterday. I get that you guys don't understand the language, but the footage show clearly what I describe above. Both the CCS-port and that it is a pure EU-version.

The US spec showroom cars that was on display in Europe since december was also the non-P version. All the EU-cars for the press that arrived last week are all Performance versions.

At about 6:00 in the video you can see the car being CCS-charged

 

[PhaseWhite]

I wonder if they wanted to optimize the news cycles or something. Maybe they do have something almost ready to announce and they want it to make all the headlines and act as another demand lever....

I suspect it has more to do with Tesla being afraid that if they announce AP3 people defer orders until it is shipping in all new cars.
The fact that Telsa did not mention one peep about AP3 on the Q4 earnings call is bullish, it means it's almost ready (basically on schedule). We probably won't hear anything until they are already shipping cars. My guess is it will be announced as shipping on all new cars during Q1 earnings report.

 

[mongo]

I suspect it has more to do with Tesla being afraid that if they announce AP3 people defer orders until it is shipping in all new cars.
The fact that Telsa did not mention one peep about AP3 on the Q4 earnings call is bullish, it means it's almost ready (basically on schedule). We probably won't hear anything until they are already shipping cars. My guess is it will be announced as shipping on all new cars during Q1 earnings report.

As soon as they run out of AP2.5 units, and not a moment before.

 

[mongo]

@M3BlueGeorgia , why the disagree? HW3 has backward compatible variants for S, X, and 3. There is no reason for Tesla to keep any 2.x stock.

 

[DrDabbles]

The thought that the HW3 components will be more expensive is wrong in two ways. First, by designing a purpose built component with their own in-house design team and then farming the fabrication out to a competent supplier means they don't pay the necessary profit margin of another company. Basically, eliminating the middle person and all of their overhead.

By having designed a purpose built device with little component waste or overhead, the actual processing units themselves are less expensive, which is yet another win. And finally, the most important metric when designing processors and systems, the cost per operation. A purpose built processor is always going to be faster than a general purpose one, so operations are already optimized for the work being done. That means more work can be done in the same amount of time (usually), which means for every dollar spent you're doubling your per-dollar efficiency. Since we've heard that HW3 is an order of magnitude faster than HW2.5, that would mean even if they units cost the same to produce, the HW3 unit is an order of magnitude more efficient for the money spent.

This all makes HW3 an extremely wise investment. Now we just have to wait to see if Tesla changes the algorithms/network they use and whether or not their chips have the flexibility to run them or not.

 

[jimmy_d]

The thought that the HW3 components will be more expensive is wrong in two ways. First, by designing a purpose built component with their own in-house design team and then farming the fabrication out to a competent supplier means they don't pay the necessary profit margin of another company. Basically, eliminating the middle person and all of their overhead.

By having designed a purpose built device with little component waste or overhead, the actual processing units themselves are less expensive, which is yet another win. And finally, the most important metric when designing processors and systems, the cost per operation. A purpose built processor is always going to be faster than a general purpose one, so operations are already optimized for the work being done. That means more work can be done in the same amount of time (usually), which means for every dollar spent you're doubling your per-dollar efficiency. Since we've heard that HW3 is an order of magnitude faster than HW2.5, that would mean even if they units cost the same to produce, the HW3 unit is an order of magnitude more efficient for the money spent.

This all makes HW3 an extremely wise investment. Now we just have to wait to see if Tesla changes the algorithms/network they use and whether or not their chips have the flexibility to run them or not.

I agree. Producing their own NN processor required Tesla to build out a dev team and infrastructure and put money into IC development for years before the IC was ready. But all that is sunk cost and much of it was years ago - especially the risk margin. Incremental cost of fabricating units today is certain to be less expensive than a comparably capable system with OTS components and likely to be less expensive than HW2.5.

Since there are no comparably capable commercial ICs available today the first comparison is probably moot. It's not possible to build HW3 today with any widely available IC. The comparison to HW2.5 depends on what performance envelope, process node, environmental requirements, die area, and packaging requirements they had for the NN processor but it would be surprising to me if the unit cost of HW3 is more than HW2.5 given Nvidia's margins and that GP106 is not well suited to the task that Tesla is using it for.

 

[fsl2019]

Hi,

in respect to the Samsung SOC I have some ideas and remarks.

First of all: I saw one youtube video where somebody claimed to have root access to the a Tesla Model 3.

Here the processor can be identified as a INTEL Apollo Lake 3950. So my first question is know:

@verygreen that Samsung SOC you are referring to, is that really located in the ADAS portion? Just want to check on that. I am pretty sure, that you are connected with debugger to the ADAS PCB.

As to Samsung: So I can imagine that TESLA is doing there own AI chip development but on only with the strong cooperation with a technology supplier. In theory TESLA could work with TSMC e.g. directly but I to be honest I don't believe there are many OEMs out there who can do that (and also Apple did gain this experience only over the last decade ...). Also Mobileye is not able to manufacture the EQ5 with TSMC without a strong technology partner currently. In the next step (EQ6) Intel will jump in ...

So coming from that perspective I had - after knowing that TESLA is moving away from the current NVIDIA architecture to something else - two suspects which potentially work with TESLA:

a) AMD (with the background of Globalfoundries)
b) SAMSUNG

So with the Equinox SOC you mentioned here and counting A+B I think it is likely that TESLA is working with SAMSUNG also on the TRIP SOC. This is still an assumption which I cannot prove but a least - imho - has some logic.

Does that make sense?

Btw. Samsung coming from the mobile computing business for sure is trying to enter also the Automotive business segment and has underlined this with the acquisition of Harman in 2017. And that could be one other puzzle piece ... the Tier1 one could be Harman.

Best regards,
Frank

 

[chillaban]

The Intel Atom SoC is for the MCU (infotainment), not for APE (autopilot). Autopilot 2.0 and 2.5 use an Nvidia 6-core SoC, with 2 half-decent Denver cores (from NVIDIA's failed attempt to create and license an x86 core, these were then repurposed to use binary translation to run armv8 instruction sets) and 4 energy efficient next-to-useless cores. That is paired with a midgrade NVIDIA GPU.

I think Samsung Exynos is a pretty no-brainer replacement for the applications processor in the current APE board. nvidia peripherals (IP blocks) tend to be kind of weird and buggy at the same time, while the Samsung ones are more sensibly designed.

The custom TRIP accelerator I think is the more exciting part. The rest of the SoC just needs to boot Linux and be able to deal with 8 camera feeds and have a decent enough PCIe implementation to meet the requirements of connecting the TRIP.

 

[Kant.Ing]

Very likely it's integrated with Samsung IPs, including that Exynos, on a single chip, and fabricated by Samsung fab.

 

[wcorey]

the TRIP chip is the Tesla chip (they still contract some fab to make it, likely Samsung sicne they switched to their SoC too?) it's not a general purpose CPU, more like a Google TPU, very special purpose device.

It's also been reported Tesla has asked the administration for an exemption from the Chinese tariff, indicating an imminent influx of V3 chips from China.

 

[kengchang]

Very likely it's integrated with Samsung IPs, including that Exynos, on a single chip, and fabricated by Samsung fab.

Doesn't make sense if Tesla wants to stay vendor neutral on the ARM CPU to go with single package SoC.

 

[mongo]

Doesn't make sense if Tesla wants to stay vendor neutral on the ARM CPU to go with single package SoC.

What vendors have pin for pin compatible CPUs?

ARM is a standard, who fabs it is mostly irrelevant especially if they are fabbing the rest of your chip.

 

[wcorey]

It's also been reported Tesla has asked the administration for an exemption from the Chinese tariff, indicating an imminent influx of V3 chips from China.

https://www.reuters.com/...tesla/tesla-urges-tariff-exemption-for-chinese-made-car-co...

 

[chillaban]

What vendors have pin for pin compatible CPUs?

ARM is a standard, who fabs it is mostly irrelevant especially if they are fabbing the rest of your chip.

Nobody. ARM is much more modeled around making core IP and it’s not their problem to give you a working system including a DRAM controller, essential peripherals, etc.

Choosing an ARM soc from anybody that runs Linux is sufficiently vendor neutral. As is not tying yourself to the nvidia Drive PX SDK. And it seems like Tesla has done all of that.

 

[kengchang]

What vendors have pin for pin compatible CPUs?

ARM is a standard, who fabs it is mostly irrelevant especially if they are fabbing the rest of your chip.

Isn't he talking about putting ARM+TRIP into a single chip? At least that's what I am reading, which I don't think Tesla will do.

Very likely it's integrated with Samsung IPs, including that Exynos, on a single chip, and fabricated by Samsung fab.