Tesla autopilot HW3

[EVNow]

What if FSD is good enough to average about 15k miles between accidents? If it has enough close calls you might still be attentive. But if it has operated successfully for year or two (with luck), how much attention will you be paying while it's driving in year 3? And that's about a 5x worse than human average accident rate, right? Will the driver be able to avoid that one time a year when FSD screws up?

I think FSD - when it is good enough - will not screw up and cause crashes. It will likely go to the side of the road and stop in unusual circumstances. Humans can handle that.

Good thing about FSD is that it won't crash because of distraction or lack of sleep. That is where the majority of better than human averages will come from.

ps : BTW, that once in 10 year rate (4.2 / million miles)is only for serious accidents. Smaller accidents are much more common (like scraping tires or hitting the curb).

 

[daniel]

AP isn't much of a test problem. We regularly have to disable it.

What if FSD is good enough to average about 15k miles between accidents? If it has enough close calls you might still be attentive. But if it has operated successfully for year or two (with luck), how much attention will you be paying while it's driving in year 3? And that's about a 5x worse than human average accident rate, right? Will the driver be able to avoid that one time a year when FSD screws up?

For many of us with repetitive drives maybe it will perform those drives like clockwork, with only strange obstacles causing any new problems. Maybe we'll only have to pay attention on new-to-us routes or when a camel steps in front of the car.

When we first get FSD features I'm sure we'll be busy working with it. But if FSD becomes almost as good as a human, the human may have problems catching the errors and getting it to human accident rates or better. Nonetheless I can't wait to try!

In the progress of development of FSD, at any particular time there will be some average rate of accidents for a hypothetical test car without a driver. The question you raise is a valid one: Will drivers be lulled into complacence when the need to intervene drops to once every two years?

I think the thing you're missing is that the kinds of mistakes computers make are very different from the kinds of mistakes people mistake, and we should not look only at the error rate of the FSD. We also need to look at the rate of errors prevented by the FSD. People are not good at paying attention. This is why AP and EAP nag you if they detect that your hands are not on the wheel. The nag will become more and more important as the error rate of the computer falls. They really cannot move to Level 3 until the accident rate of the computer is well below the accident rate of humans.

Your concern about drivers becoming complacent might mean that the nag has to get more and more obnoxious until FSD is truly ready to operate on its own.

 

[CK_Stuggi]

Depending on the reliability of AP I agree with @tomc603 . As @daniel illustrates, he thinks "he knows" where AP will fail. Even Elon has stated somewhere that currently the problem are people overestimating the abilities of AP (nb - that was last year or even the year before, around the time of the Mountain View crash I think).

That's the exact weak point of Tesla, but then again I'm sure the other aspect also applies; That of rolling out safety features asap even if they're not 100%, to allow for the saving of lives even during maturation.

Yes, there were the three fatal crashes; but we dont have any idea how many lives AP in its progressing development has already saved.

So, even accounting for people who overestimate APs ability, I still consider the "L2 supervising safety driver" a pretty smart approach.

 

[DrDabbles]

Yeah, @CK_Stuggi, I think one of the things I've said here in the past is that there's features that could be extremely useful, that are part of what will be needed for FSD, and likely have a major safety advantage. Adding stop sign and red light braking requires recognizing them. That unlocks at least the ability to alert the driver and engage EAB if necessary, even before AP can use the feature.

 

[cutesheep]

Hi every automated vehicles fans,
I really need your help and input regarding my dissertation! I would really appreciate it if you could feel out my online survey regarding the features of automated vehicles in the online community. (Im postgraduate of University of Warwick)

Please click on the link below.. it shouldn't take more than 3 mins to complete it. Thank you!!!!!!
https://warwickwmg.eu.qualtrics.com/jfe/form/SV_7WdAPlhRJOSLDjT

 

[joefee]

Just a heads up. Regular service personnel do not know what's going on internally at Tesla and any prediction they give you regarding the hardware upgrade, or anything to do with autopilot development in general, is likely a complete guess and wrong.

Most employees stay informed like the rest of us.....with Elon's tweets/50% discount for over optimism!

Take this poll so we can keep track of the FSD upgrade progress:

HW3 upgrade waiting room: All FSD Tesla’s

 

[diplomat33]

Tesla confirms in the Q2 update letter that AP3 computer is in production since Q2:
Our new FSD computer, which is ~21x faster than our previous HW2.5 computer, has been in production since Q2,” Tesla wrote.
Tesla Autopilot's stop sign, traffic light recognition and response is operating in 'Shadow Mode'

 

[DrDabbles]

which is ~21x faster than our previous HW2.5 computer

That's ~ 21x*

*21x speed includes performance from both redundant FSD chips, so actual speed increase is 10x

 

[EVNow]

That's ~ 21x*

*21x speed includes performance from both redundant FSD chips, so actual speed increase is 10x

Hmm ... Both can't be redundant.

 

[heltok]

But do they need to very that every frame at 100fps is identical? Lets say check that both cores are in agreement every second and thus get a 99% increase in performance is that enough?

 

[mongo]

Hmm ... Both can't be redundant.

It only takes one redundant chip to drop the total unique processing in half.
(Yes phrasing, two chips operating in a redundant set up, RAID 1, if you will)

But do they need to very that every frame at 100fps is identical? Lets say check that both cores are in agreement every second and thus get a 99% increase in performance is that enough?

I few issues: at highway speed 1 second is 88 feet, more than half the stopping distance. Changing the processing time every X time units causes jitter in the sample-action response delay which can mess up control loops.
Much simpler and safer to just do everything on both and cross check on every action (or even instruction). Dual lock step cores is the standard setup for safety critical processing.

 

[rnortman]

Do we have any reports about performance/stability of the HW3 cars that have been produced so far?

 

[Saghost]

Do we have any reports about performance/stability of the HW3 cars that have been produced so far?

I thought we were told that AP3 is currently running slightly modified AP2.5 code, rather than anything developed to take advantage of that hardware.

My Raven X does AP better than my AP1 car did, and hasn't rebooted for an AP computer freeze yet, but neither of those seems like a useful answer, and I'm not sure any answer is meaningful until they get their own codebase.

 

[rnortman]

I thought we were told that AP3 is currently running slightly modified AP2.5 code, rather than anything developed to take advantage of that hardware.

But the question is -- is it actually stable and getting the same performance when running the "same" software? I highly doubt it is truly the same; they may have ported the same network and same weights from the Nvidia GPU to their new chip, but the chip works differently so the implementation is sure to be different. And the code surrounding that chip (the stuff running on normal CPU cores) must also have changed in some ways. And all else being equal, the new chip should run the same network faster, which may also have effects (maybe good, maybe bad).

When AP2.5 first came out, it was super buggy, despite being the "same" software. I know, because I was one of the first with AP2.5.

 

[Saghost]

But the question is -- is it actually stable and getting the same performance when running the "same" software? I highly doubt it is truly the same; they may have ported the same network and same weights from the Nvidia GPU to their new chip, but the chip works differently so the implementation is sure to be different. And the code surrounding that chip (the stuff running on normal CPU cores) must also have changed in some ways. And all else being equal, the new chip should run the same network faster, which may also have effects (maybe good, maybe bad).

When AP2.5 first came out, it was super buggy, despite being the "same" software. I know, because I was one of the first with AP2.5.

I think the major point is the neural networks - if I remember right, AP3 can handle much larger matrixes going into the NN than AP2.5. But I could be confused.

Certainly there are some differences in the code already - AP3 records videos in H265 with hardware support, vs the H264 of earlier versions.

I'm not sure how we'd get a proper objective answer to your question. Anecdotally, my new X seems to be doing at least as well as what I read on the forum, possibly better (very few phantom braking events, seems smoother than described, no spontaneous rebooting,) but that's not really a definitive answer.

We'd really need to have rooted versions of both computers on a bench we could run cases against to have a proper answer.

 

[emmz0r]

Very interesting video indeed.
This guy says that HW3 actually performs better on the exact same route. He says it feels like as in a game, where HW3 is "max settings" and HW2 is "medium settings" when it comes to quality.

I think Elon has a white lie when it performs worse / same as HW2. It has to at least run in a better rate...

 

[diplomat33]

I think Elon has a white lie when it performs worse / same as HW2. It has to at least run in a better rate...

it would stand to reason that Elon might have been downplaying the abilities of the FSD computer compared to AP2 to avoid pandemonium at delivery centers, with owners demanding the new FSD computer upgrade. That is why he said that owners should wait for the FSD software before asking for the FSD computer because they would not see any improvements before then.

 

[emmz0r]

it would stand to reason that Elon might have been downplaying the abilities of the FSD computer compared to AP2 to avoid pandemonium at delivery centers, with owners demanding the new FSD computer upgrade. That is why he said that owners should wait for the FSD software before asking for the FSD computer because they would not see any improvements before then.

That brings me to another point. Isn't it better to just do retrofits in a steady pace instead of having a huge wave of them and a ton of pissed off customers when FSD finally does come? Maybe internally they think they are way off still...

 

[diplomat33]

That brings me to another point. Isn't it better to just do retrofits in a steady pace instead of having a huge wave of them and a ton of pissed off customers when FSD finally does come?

Yes and my assumption is that Tesla will be doing a gradual upgrade schedule over time.

 

[emmz0r]

Yes and my assumption is that Tesla will be doing a gradual upgrade schedule over time.

So preferably starting before FSD is rolled out which is end of this year. Boy, they are in a rush then