Tesla autopilot HW3

[boonedocks]

Then, at the risk of sounding what we in the UK call a bell end, it is not Tesla's fault. Its the fault of the driver/s that went to sleep.

People fall aslepp at the wheel in all sorts of cars. They die asleep at the wheel in all sorts of cars. The fact that a Tesla makes a system which makes driving a less intense activity and therefore gave them, and you it would seem, a poor excuse to fall asleep, does not exhonerate them from blame.

If you feel sleepy behind the wheel you pull over, whether you are driving a Tesla or a tractor.

Yes but none of those make the headline news. Conveniently your argument won't fit his narrative.

 

[RedBall]

Once other manufacturers start deploying the same level of tech, we shall see who has been swimming naked.

My bet is on the humans, not the tech.

 

[mongo]

Source? Got board pics you can PM me? The new neural net isn't in use, I call BS, it is on dev cars somewhere...
And if some kind soul that wants 5 mins of fame and has a torx screwdriver that got an APH4 delivery will PM me, we can stop speculating

Yeah, Telsa employees are likely running the new NN.
Regarding HW3, peeps have dissected the binaries...
Tesla autopilot HW3

The “TRIP” device obviously is the most interesting one. A special firmware that encompasses binary NN (neural net) data is loaded there and then eventually queried by the car vision code. The device runs at 400MHz. Both “TRIP” devices currently load the same NNs, but possibly only a subset is executed on each?

With the Exynos SoC being a 2015 vintage and in consideration of comments made by Peter Bannon on the Q2 2018 earnings call, (he said “three years ago when I joined Tesla we did a survey of all of the solutions” = 2nd half of 2015), does this look like the current HW2/HW2.5 NVIDIA autopilot units were always viewed as a stop-gap and hence the lack of perceived computation power everybody was accusing Tesla of at the time of AP2 release was not viewed as important by Tesla?

 

[Silicon Desert]

Yes but none of those make the headline news. Conveniently your argument won't fit his narrative.

Well I don't know. They seem to fall asleep on tractors at all ages

 

[Silicon Desert]

Every time a Tesla crashes the first question is was AP engaged? /QUOTE]

Yea, that's an interesting point. It does seem like every time there is a Tesla accident, that is first question. When there is a deadly accident in an ICE car, they say the driver must have been drunk. I guess it is assumed by the media that there are never any Tesla drivers under the influence

 

[SR22pilot]

If current HW3 installations are using the NN processor then I stand corrected. We should know soon enough. I suspect there will soon be people who have one car with HW2.5 and another with HW3. If there is a dramatic performance difference then I stand corrected. There is certainly a large chance that I am wrong.

 

[MorrisonHiker]

If current HW3 installations are using the NN processor then I stand corrected. We should know soon enough. I suspect there will soon be people who have one car with HW2.5 and another with HW3. If there is a dramatic performance difference then I stand corrected. There is certainly a large chance that I am wrong.

Soon, some will have AP2, AP2.5 and AP3.

 

[SR22pilot]

No ADAS includes level 2.



There has absolutely been zero deaths because other car manufacturers weren't gross negligent.

Again I repeat, There are probably around 100 million Level 2 ADAS. All of them except super cruise (which uses camera) has a 7-15 secs timer and most don't use the stupid torque sensor that Tesla but touch capacity sensor aswell. Also and more importantly the CEOs doesn't hype the product up to be self driving.

This is in contrast to Tesla who started with AP 7.0 or 7.1 and allowed you sometimes upwards of 30 mins to 1hr without placing hands on the wheel. Immediately 2 people died from that. They then decreased it but not enough and still more people died.

• Do you think if the initial 30 mins were still here would there be more death count or less?
• 
  
• Do you think if it was 7-15 seconds as other car manufacturers with capacitiy touch that there would be ZERO deaths.
• 
  
• Do you think if other manufacturers allowed upwards of 30 without placing hands and no way to know you are even paying attention in those 30 minutes, that there won't be deaths?

Remember that all deaths happened because driver fell asleep trusting the system. As Elon said himself its experienced AP drivers that die not inexperienced ones that don't trust the system.

The torque sensor, which I find a pain, was added because idiots posted videos of themselves getting in the back seat of the car while the car was on AP. As soon as I saw that I remarked to a friend that idiots would end a good thing and they did.

For the record, I love what Cadillac has done with recognizing when you are paying attention and I wish Tesla did the same. The Cadillac system is much less annoying.

Please point me to a reliable source of crash data for other manufacturers that shows the mode of the car at the time of the crash. Can't do it? Then stop posting personal opinion as fact. I personally think many have died using the other systems but I will admit that I can't prove it. There are indicators such as a French study (Cruise Control Could Be Hazardous To Your Health) but I don't see data on the other systems.

 

[DoubleE]

I test drove a brand new Model X P100DL, and used the drive on autopilot - I would say it drove extremely smoothly. Not sure how it would compare to H2.5, but I'm highly considering trading up now.

 

[MorrisonHiker]

I test drove a brand new Model X P100DL, and used the drive on autopilot - I would say it drove extremely smoothly. Not sure how it would compare to H2.5, but I'm highly considering trading up now.

Can you tell us how you know it had AP3? The vast majority of cars being delivered still have AP2.5 hardware.

 

[Zaxxon]

Again I repeat, There are probably around 100 million Level 2 ADAS. All of them except super cruise (which uses camera) has a 7-15 secs timer and most don't use the stupid torque sensor that Tesla but touch capacity sensor aswell.

Since you keep talking without knowing about what you speak, I'll add a correction here. We have a 2018 Volvo with PilotAssist II, which uses a torque sensor and no touch sensor. It also frequently waits far longer than 15 seconds to nag, does not give any audible indication when it decides to disengage, and would happily drive me off of a cliff if I failed to pay attention--much more easily than Autopilot would. To say it's a safer system than Autopilot is, to use a Tesla term, ludicrous.

 

[stretchsje]

I went back and read the first ~10 pages of this thread. Good stuff.

So the Tesla Neural Net processor runs as a PCI-X device, but HW 2.5 does not have a PCI-X slot with which to add it. Thus the entire SoC is going to be replaced for FSD vehicles, which in the model 3 is a reasonably quick process by design. In doing so, the nVidia SoC will be replaced with a Samsung SoC. Assumptions here seem to be that the new SoC gains two high-speed CPU cores in the process, but that the GPU is probably less impressive (and unnecessary, though I think that could be disputed- isn't sensor fusion and telemetry done efficiently in the GPU?)

The new HW3 seems to have an A and B operating mode, but I'm unclear if these are strictly logical modes. Elon has mentioned the NN accelerater having enough processing power leftover to have "redundancy" but I'm unclear on what that means either. Spare processor cycles != redundancy! And in the case of neural networks, I don't think there is such a thing as too much processing power.

Some are speculating that HW3 cars currently shipping are running the existing HW2.5 NN, but the software stack would be quite different. Is HW2.5's NN written in nVidia's Drive API or CUDA, or is it in a language (OpenCL?) portable to the new Samsung hardware?

I doubt the practice of putting unused hardware into the vehicles will continue as operating margins become increasingly important with lower average sale prices. This makes me wonder if non-FSD cars in the future will continue to run the older neural net in the GPU, to be used only for advanced cruise control, emergency braking, and such. This could be the reason for the software-defined "A" and "B" modes.

 

[mongo]

I went back and read the first ~10 pages of this thread. Good stuff.

So the Tesla Neural Net processor runs as a PCI-X device, but HW 2.5 does not have a PCI-X slot with which to add it. Thus the entire SoC is going to be replaced for FSD vehicles, which in the model 3 is a reasonably quick process by design. In doing so, the nVidia SoC will be replaced with a Samsung SoC. Assumptions here seem to be that the new SoC gains two high-speed CPU cores in the process, but that the GPU is probably less impressive (and unnecessary, though I think that could be disputed- isn't sensor fusion and telemetry done efficiently in the GPU?)

The new HW3 seems to have an A and B operating mode, but I'm unclear if these are strictly logical modes. Elon has mentioned the NN accelerater having enough processing power leftover to have "redundancy" but I'm unclear on what that means either. Spare processor cycles != redundancy! And in the case of neural networks, I don't think there is such a thing as too much processing power.

Some are speculating that HW3 cars currently shipping are running the existing HW2.5 NN, but the software stack would be quite different. Is HW2.5's NN written in nVidia's Drive API or CUDA, or is it in a language (OpenCL?) portable to the new Samsung hardware?

I doubt the practice of putting unused hardware into the vehicles will continue as operating margins become increasingly important with lower average sale prices. This makes me wonder if non-FSD cars in the future will continue to run the older neural net in the GPU, to be used only for advanced cruise control, emergency braking, and such. This could be the reason for the software-defined "A" and "B" modes.

Might just be terminology, but the whole module gets swapped, not just an SoC. New chip has a Tesla designed custom processor for NN, not a GPU (not design by Samsung either).

 

[lunitiks]

And if some kind soul that wants 5 mins of fame and has a torx screwdriver that got an APH4 delivery will PM me, we can stop speculating

Ney, that soul will not PM you, it will PM me.
And it shall not be granted 5 mins of fame, but an everlasting testament - right here, on the TMC - for being the bravest of them all; "them" being the first wave of HW3 owners 2019.

And it shall not be a torx screwdriver, but a torx screwdriver and some plastic pry bar tools.

 

[stretchsje]

Might just be terminology, but the whole module gets swapped, not just an SoC. New chip has a Tesla designed custom processor for NN, not a GPU (not design by Samsung either).

Right. Old: nVidia CPU + GPU. New: Samsung CPU + GPU + Tesla NN accelerator. Likely a very different PCB and heat dispersion. I was using that terminology to differentiate that there is both a new SoC AND a Tesla NN accelerator as part of HW3.

 

[mongo]

Right. Old: nVidia CPU + GPU. New: Samsung CPU + GPU + Tesla NN accelerator. Likely a very different PCB and heat dispersion. I was using that terminology to differentiate that there is both a new SoC AND a Tesla NN accelerator as part of HW3.

Gotcha, I think the GPU is just a built in part of the Samsung processor core as opposed to an intended NN processor. The A vs B is, I believe dual TRIP (NN processor) cores for safety redundancy.

 

[S4WRXTTCS]

The torque sensor, which I find a pain, was added because idiots posted videos of themselves getting in the back seat of the car while the car was on AP. As soon as I saw that I remarked to a friend that idiots would end a good thing and they did.

For the record, I love what Cadillac has done with recognizing when you are paying attention and I wish Tesla did the same. The Cadillac system is much less annoying.

Please point me to a reliable source of crash data for other manufacturers that shows the mode of the car at the time of the crash. Can't do it? Then stop posting personal opinion as fact. I personally think many have died using the other systems but I will admit that I can't prove it. There are indicators such as a French study (Cruise Control Could Be Hazardous To Your Health) but I don't see data on the other systems.

The torque sensor was in there from the beginning, and it was actually an MB part as the Model S used a few parts from the MB.

The first "getting in the back seat" videos weren't from Tesla vehicles at all. It's funny how people keep thinking they only started when Tesla AP1 came around.

WK057 (a user on TMC) was actually the first person in a Tesla to do the "in the back seat" thing. He got the idea from other youtube videos out there with other cars. I think it was actually an Infinity that was his inspiration if memory serves correct.

In any case he was the first, and did so in a very safe/sane manner. It wasn't some idiotic thing, but just to prove it could be done. After those videos Tesla got the bright idea to use the seat sensor to prevent that. Of course people could just put weight down on the seat.

Tesla didn't change the behavior of the torque sensor until the first fatality AP related accidents happened in the US.

They went away from a "confidence" based warning to timed nags (that are also speed limited related).

Blader can't provide any statistics because no one other than Tesla keeps them.

Even adaptive cruise control can kill a person if a person doesn't understand that a radar based ADAS system can't always see stopped objects. There is absolutely no way that people haven't died while using adaptive cruise control (which counts at a SAE L1 system). All a person has to do is look away at the wrong time. The average person that is going to txt while driving will use anything the car has as a way to buy time to send that txt.

If they have an adaptive cruise control system they'll use that.
If they have a L2 system they'll use that
If they have to have a hand on the steering wheel then they'll have one hand on the steering wheel while they txt.

When it comes to texting the steering wheel nag doesn't really matter.

The only GOOD way of implementing an ADAS system is how Cadillac did it. I don't think there is any disagreement on that one. No one wants the stupid nag anymore.

Eye Tracking is the only way to save the texters from themselves, and to deal with people who fall asleep while driving.

There are some specific things about Tesla.

You can simply drive wicked fast. This is the number one killer when it comes to Tesla. It's getting to the point where it's every month while we stand around talking about one AP death every year or two.
The AP system is good enough to trick you into trusting it. Then it throw you into a gore point and straight for a divider.
It won't geofence anything so its not going to protect stupid people from themselves
The Model S/X are especially quite and smooth. I don't fall asleep when I drive, but all my passengers are sound asleep even after only 30 min. Or maybe it's because I'm really boring.

But, there is also one important thing that isn't unique to Tesla.

There is a loss of situational awareness when driving any L2 system. This can lead to accidents because the user is unprepared mentally to take over. Part of why I stopped using AP was because of this. Nothing about the nag system will impact that. It's simply a weakness I discovered in myself (that some others have as well). I proved to myself that I was a worse driver with AP because my reaction time to debris on the road was worse with AP than without. When L2/L3 cars have Lidar, rear corner radars, and some other tech then I'll use L2/L3 technology a lot more. Until then it's just something to play with.

 

[verygreen]

Thus the entire SoC is going to be replaced for FSD vehicles, which in the model 3 is a reasonably quick process by design

no. you have it wrong. It's a quick process on S/X with hw2.5. It's long and painful on model3. really long (multiple hours) and really painful (draining the coolant, taking a bunch of stuff apart inside the cabin and outside)

 

[vinny84]

Interesting. Did you check the firmware on the car? And did Tesla tell you the car had hw3?

I test drove a brand new Model X P100DL, and used the drive on autopilot - I would say it drove extremely smoothly. Not sure how it would compare to H2.5, but I'm highly considering trading up now.

 

[Bladerskb]

Since you keep talking without knowing about what you speak, I'll add a correction here. We have a 2018 Volvo with PilotAssist II, which uses a torque sensor and no touch sensor. It also frequently waits far longer than 15 seconds to nag, does not give any audible indication when it decides to disengage, and would happily drive me off of a cliff if I failed to pay attention--much more easily than Autopilot would. To say it's a safer system than Autopilot is, to use a Tesla term, ludicrous.

Looks like i you missed the fact that i said "most" and Volvo's User guide states 10-15 secs, secondly the very fact that you don't trust it proves its safer than AP. Remember what Elon said, people die and get into accidents with AP because they TRUST the system, not because they DON'T trust it.

There's like dozens of videos of people sleeping on AP while going 70 mph+. Theres one every other week.