How does Tesla know 20x more compute than HW 2.5 is enough?

[Atari2600]

No one seems to question if 20x more compute than HW 2.5 is enough for FSD to be reliable and fully functional. Elon thought HW 2.5 would be enough and it was not. Is there any metric being used to show HW 3 is powerful enough to satisfy people like Edward Markey with a true AutoPilot solution? If history repeats itself I think this will be devastating to Tesla(I don't want this to be the case).

 

[jmaddr]

Compute is never enough assuming time goes by. Eventually, it's capabilities will be overrun by what is possible. So the definition of FSD is what's important here.

Is FSD Level 3...where a human still needs to be in a drivers seat, ready to take over but it can navigate anywhere you tell it? I personally think this would be possible with HW3 and what Elon is thinking. Or is FSD more like Level 4...the first level of total system control but there would be some situations like environmental that would not be supported. This would be a likely maybe in my opinion. Or does it mean level 5, whereby the drivers seat essentially becomes a passenger seat and the steering wheel and pedals can be removed from the car and all environmental conditions are supported. I would think HW3 would be pushed to sustain L5 but that's a random internet opinion so please take it just like that.

 

[SSonnentag]

They know it's enough the same way they knew v2.0 hardware was enough.

 

[Atari2600]

I thought the original definition of FSD was at level 5. I realize Tesla rewrote the definitions after I purchased my car.

Options
Enhanced Autopilot
Full Self-Driving Capability

EAP included most of what Autopilot and FSD includes today. After reviewing the old definition I changed my statement of expectation to level 5.

 

[Atari2600]

FYI this what I purchased for reference.

 

[boonedocks]

FYI this what I purchased for reference.
View attachment 504758

You and I and so many others completely bought into that as well. (Me....3 times now)

 

[Atari2600]

You and I and so many others completely bought into that as well. (Me....3 times now)

I hope you win three prizes!
The definition of FSD is now drastically different than 2018 and earlier. Our version is pretty clear cut expectation wise. I should be able to sleep at the wheel.

Do you have concerns about HW 3 being powerful enough?

 

[Eno Deb]

No one seems to question if 20x more compute than HW 2.5 is enough for FSD to be reliable and fully functional.

There are plenty of people who question this.

Elon thought HW 2.5 would be enough and it was not. Is there any metric being used to show HW 3 is powerful enough to satisfy people like Edward Markey with a true AutoPilot solution?

I don't think Tesla knows either. HW3 is just what's currently available given the state of technology, cost constraints and the development time they had. It also appears to be highly specialized for a specific way of doing things, which always brings the risk that the software could develop in a different direction and make the hardware less suitable.

 

[VT_EE]

I don't know why people keep saying level 5 means it should operate in ALL conditions. There are plenty of conditions where humans shouldn't be operating the vehicle either. Take an ice or heavy snow storm. No way is any autonomous system going to be safe driving in those. It has to do with physics, not how good your AI is.

 

[Atari2600]

I don't know why people keep saying level 5 means it should operate in ALL conditions. There are plenty of conditions where humans shouldn't be operating the vehicle either. Take an ice or heavy snow storm. No way is any autonomous system going to be safe driving in those. It has to do with physics, not how good your AI is.

I'll be happy as long as it conducts trips with no action required by the person in the driver's seat.

 

[mswlogo]

My guess is Sensors (lack of) will be the next limitation, not HW3.

It needs Radar in the rear corners (like most blind spot cars have).
The cameras need to be hermetically sealed or heaters.
The cameras need washers.
The cameras need to have better dynamic range (sun behind a traffic light completely blinding function is a non starter).
Not sure how to solve a truck in front of you completely blocking traffic lights.
Front camera needs to "see" further ahead.

Like mentioned in other threads, to be feasible, I think all cars need to talk to each other.
Simple example, multiple cars arriving simultaneously at a 4 way stop. Who goes first? This is even awkward for human's sometimes. No you go, No you go, No you go. If she's cute, I let her go first

I think we will just end up with smarter and safer "assist" until a major do over of sensors.

 

[Atari2600]

My guess is Sensors (lack of) will be the next limitation, not HW3.

It needs Radar in the rear corners (like most blind spot cars have).
The cameras need to be hermetically sealed or heaters.
The cameras need washers.
The cameras need to have better dynamic range (sun behind a traffic light completely blinding function is a non starter).
Not sure how to solve a truck in front of you completely blocking traffic lights.
Front camera needs to "see" further ahead.

Like mentioned in other threads, to be feasible, I think all cars need to talk to each other.
Simple example, multiple cars arriving simultaneously at a 4 way stop. Who goes first? This is even awkward for human's sometimes. No you go, No you go, No you go. If she's cute, I let her go first

I think we will just end up with smarter and safer "assist" until a major do over of sensors.

Thanks, yes there are sensor questions which are often brought up in threads and I fully agree... but I am specifically concerned with the processing power here. If the processing power isn't enough then the whole sensor discussion does not matter.

But to support your points I enjoyed the follow message in the middle of a 115 mile trip last Friday:

 

[diplomat33]

No one seems to question if 20x more compute than HW 2.5 is enough for FSD to be reliable and fully functional. Elon thought HW 2.5 would be enough and it was not. Is there any metric being used to show HW 3 is powerful enough to satisfy people like Edward Markey with a true AutoPilot solution? If history repeats itself I think this will be devastating to Tesla(I don't want this to be the case).

I don't think anyone outside of Tesla really knows. Presumably, Tesla has some metric that they are using. They are probably comparing the size of the NN that they think is big enough to handle all the tasks for autonomous driving with how big of a NN the chip can handle. I think I read somewhere that the current NOA is only using 20% of the compute power of AP3. If true, then Tesla may feel that AP3 has enough margin that it will be good enough to do FSD.

Also remember that Tesla is making some basic assumptions: They are assuming that camera vision alone can do FSD. And they are assuming that the current sensor set up is sufficient for FSD. So they are looking at the size of NN that can handle the current input from all the cameras as capable of doing FSD. If it turns out that FSD requires more input data (like from higher resolution cameras, additional cameras, additional radar or lidar) then that changes the size of the NN which changes the whole calculation. If FSD requires a bigger NN, then the AP3 chip may not be good enough.

Not sure how to solve a truck in front of you completely blocking traffic lights.

HD maps are really useful in these cases because the HD map can tell the car that a traffic light should be there even when the cameras can't see it.

 

[Soda Popinski]

One obvious way to do it would be to see how much you need to slow down sensor input to get "real-time" output. This would of course require some sort of functioning FSD software build. Perhaps they run what they have, and put in a fudge factor to estimate what they think they'll need for "feature complete."

Here's a wishful thinking theory I just came up with. If you've ever seen early game play footage from an upcoming video game that looks amazing compared to what eventually gets released, that's because it's game was running on super high end computers. In order to make minimum system specs, and even recommended specs run at a decent frame rate, they need to optimize the code. Make it more efficient, and drop bells and whistles that take up a lot of compute.

Put that in the context of Tesla's FSD. What if they've got a build that works now, but requires even more compute than is available in HW3. They need to optimize it to work. That could explain the acquisition of DeepScale.

 

[Uncle Paul]

Imagine that V3 computer will be good enough for FSD, but that Tesla is working on much faster and more capable chips for the future. They are working on "the next generation" of processors all the time, and I imagine they will continue to roll out advancements as they become available to make the Autopilot consistantly better.

It is typical for software writers to constantly desire more speed. They write more sophisticated software that needs increasingly faster hardware to operate optimumly. They are also working, at the same time, to make chips smaller and use less energy.

If buyers want to wait until the next generation chips roll out they will be waiting forever.

 

[mswlogo]

One obvious way to do it would be to see how much you need to slow down sensor input to get "real-time" output. This would of course require some sort of functioning FSD software build. Perhaps they run what they have, and put in a fudge factor to estimate what they think they'll need for "feature complete."

Here's a wishful thinking theory I just came up with. If you've ever seen early game play footage from an upcoming video game that looks amazing compared to what eventually gets released, that's because it's game was running on super high end computers. In order to make minimum system specs, and even recommended specs run at a decent frame rate, they need to optimize the code. Make it more efficient, and drop bells and whistles that take up a lot of compute.

Put that in the context of Tesla's FSD. What if they've got a build that works now, but requires even more compute than is available in HW3. They need to optimize it to work. That could explain the acquisition of DeepScale.

Sure, demo's are probably done on very high end systems ($12K graphics cards). But my guess is those high end systems are available to end users and the software would scale the detail accordingly. Detail and Smoothness == MIPS. I'm sure some allow you to set detail (trading FPS) and others do it based on the hardware it sees.

 

[willow_hiller]

Imagine that V3 computer will be good enough for FSD, but that Tesla is working on much faster and more capable chips for the future. They are working on "the next generation" of processors all the time, and I imagine they will continue to roll out advancements as they become available to make the Autopilot consistantly better.

HW3 is a case of maximizing processing power, minimizing electricity usage, and holding board-size constant. The major constraint was making sure it would fit in the same footprint as the HW2.5 board.

In the future, footprint and electricity may not be as big of a constraint as they are now.

 

[mswlogo]

Imagine that V3 computer will be good enough for FSD, but that Tesla is working on much faster and more capable chips for the future. They are working on "the next generation" of processors all the time, and I imagine they will continue to roll out advancements as they become available to make the Autopilot consistantly better.

It is typical for software writers to constantly desire more speed. They write more sophisticated software that needs increasingly faster hardware to operate optimumly. They are also working, at the same time, to make chips smaller and use less energy.

If buyers want to wait until the next generation chips roll out they will be waiting forever.

I know HW4 has some target specs, but my guess is HW4 is more focused on getting the wattage down (which also implies cost) than getting the MIPS up. Just a wild ass hunch.

 

[Mardak]

I would guess HW2.5 has been running near its limits over the last many months, but it has been improving over time because a better trained network of a given size improves its weights and accuracy without increasing processing loads. So if you consider HW2.5 as some base level of processing capable of providing today's quality of Highway Autopilot, doubling the size of the neural network can improve accuracy with minimal code changes. Even with a doubled processing requirement for an improved Highway Autopilot, there's still 10x capacity left to do other processing.

Can you come up with 10 (improved-)Highway-Autopilot-quality features that you would want the car to take care of?

 

[Uncle Paul]

Believe Elon said something to the effect that 2.0/2.5 were capable of fully running current FSD hardware. Said that there currently was no benefit of switching to V3, as it has not yet been optimized and can even degrade some operations.

V3 will come into it's own as additional features (better summons) are released and the speed becomes relevant.

Currently some reports are that it renders quicker, but no reports that it enables additional FSD features or performance.

Kind of the same thing going on with the newer V3 Superchargers. Only little overall speed benefits for many of the legacy cars, but might give much more substantial charging speed increases with newer releases.