-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
The new FSD computer vs HW2.0/2.5 with ONLY AP or EAP
- Thread starter insaneoctane
- Start date
diamond.g
Active Member
Knightshade
Well-Known Member
what things does AP (the version sold today- which is what the person you're replying to asked about, and which does not change lanes or attempt to take exit ramps) run into, and why do you think HW3 would help with that?
Autosteer and TACC are a part of EAP, FSD, and AP. There are at least a handful of documented incidents in which Autosteer/TACC have run into things. Those features are, I think we would all agree, by no means perfected yet.
I highly doubt that Tesla would want to chance more accidents/bad PR by not putting Autosteer and TACC on the new NN, whether those two features are bundled under AP or FSD. And it wouldn’t make sense to have two separate NNs running on the same HW to accomplish essentially the same thing (I.e. Autosteer and TACC). The car needs to recognize objects even if just tootling along in one lane, and the speed and accuracy at which it recognizes those objects will help prevent Autosteer/TACC from running into things like fire trucks or dividers.
I highly doubt that Tesla would want to chance more accidents/bad PR by not putting Autosteer and TACC on the new NN, whether those two features are bundled under AP or FSD. And it wouldn’t make sense to have two separate NNs running on the same HW to accomplish essentially the same thing (I.e. Autosteer and TACC). The car needs to recognize objects even if just tootling along in one lane, and the speed and accuracy at which it recognizes those objects will help prevent Autosteer/TACC from running into things like fire trucks or dividers.
Last edited:
My best guess is that the situation is as follows. Of course I know nothing, but that has never stopped me before.
1) Tesla is training a new neural net that includes support for lots of new FSD features. This was demoed on Autonomy Day, but has not yet been released to the field. This new net will be an HW3 replacement for the one that is now running both AP and EAP on HW2.5 and below.
2) Cars with HW3 are currently running the released NN like everyone else, but the hardware is much faster. So this *might* translate to smoother performance with the existing AP/EAP functionality. It would be interesting to compare the behavior of an HW2.5 car to an HW3 car through a difficult route to see if there is a visible improvement with HW3.
3) When the new NN is released to HW3 cars, it will provide all the same functionality as is currently available (AP/EAP). Probably it will already perform those functions better than the current NN because it will be 'aware' of more conditions. For example emergency braking may be better, etc.
4) It is unclear whether Tesla would continue to train the neural net for HW2.5 and below once the new HW3 net is released. They might, because they are continuing to curate and extend their training data, and performing additional training against the current net is probably not terrible expensive if they already have the training data. But this is highly speculative. FWIW, my sense is that the current AP2 computer is capable of much better performance than we currently have, given a better NN.
1) Tesla is training a new neural net that includes support for lots of new FSD features. This was demoed on Autonomy Day, but has not yet been released to the field. This new net will be an HW3 replacement for the one that is now running both AP and EAP on HW2.5 and below.
2) Cars with HW3 are currently running the released NN like everyone else, but the hardware is much faster. So this *might* translate to smoother performance with the existing AP/EAP functionality. It would be interesting to compare the behavior of an HW2.5 car to an HW3 car through a difficult route to see if there is a visible improvement with HW3.
3) When the new NN is released to HW3 cars, it will provide all the same functionality as is currently available (AP/EAP). Probably it will already perform those functions better than the current NN because it will be 'aware' of more conditions. For example emergency braking may be better, etc.
4) It is unclear whether Tesla would continue to train the neural net for HW2.5 and below once the new HW3 net is released. They might, because they are continuing to curate and extend their training data, and performing additional training against the current net is probably not terrible expensive if they already have the training data. But this is highly speculative. FWIW, my sense is that the current AP2 computer is capable of much better performance than we currently have, given a better NN.
Knightshade
Well-Known Member
When taking exit ramps or hitting things on local roads with cross-traffic (where no version of AP is supposed to be used at all ) though.
All things that the new AP explicitly does not support in the first place.
The fire-truck was an AP1 car, so totally irrelevant to the discussion.
Things that I suspect HW3 will do better vs HW2.5 while using AP:
- Identifying and maneuvers relative to non-car hazards (ie potholes, bikes, pedestrians, etc)
- Less/elimination of Phantom braking?
- Local roads (not currently recommended, but also not disallowed)
HW3 vehicles should be better even on regular AP or even non-AP safety functions if for no other reason than that they will have the computing power to run the NNs on full-scale images, instead of downscaling them to 1/4 the resolution (1/2 in each direction) before processing them. So they should do a better job of recognizing things, especially at a distance.
However, I don't believe they've started to do this yet, for now they're just recompiling the same NN from HW2/2.5 for HW3, with the same resolution restrictions.
However, I don't believe they've started to do this yet, for now they're just recompiling the same NN from HW2/2.5 for HW3, with the same resolution restrictions.
Knightshade
Well-Known Member
But that Model X also, besides using a much older version of the 2.x NN that has already seen a ton of improvement, crashed on an exit ramp
which is something the new AP is never intended to be used on....and as noted- The driver had received several visual and one audible hands-on warning earlier in the drive and the driver’s hands were not detected on the wheel for six seconds prior to the collision- so much like the few other crashes on some version of AP- driver error for not supervising the system as directed.
Since AP doesn't need all the cameras (since it only functions in a single lane) the 2.5 computer could already handle everything at full frame rate... (remember when EAP originally said it'd only use 4 cameras? That's because HW2.5 can handle 4 cameras no problem at full rate...and if the NN doesn't have to worry about lane keeping and TACC and changing lanes and taking exit ramps all at the same time, you don't ever need more than 4 cameras.
The new AP is explicitly not intended for these roads- so seems unlikely they'd make it a lot better there- use on local roads will be a major selling point of the FSD upgrade so they're not going to torpedo their own sales like that.
No, that doesn't logically follow, for me. The actions for AP/EAP are within tightly constrained bounds. (Such as, behind a vehicle & in a lane.) The NN enhancements matter more for more complex and unbounded scenarios, which are encountered during freer driving, off-highway, etc.
The code for AP/EAP may not benefit from the better hardware at all, it depends on whether they do a new implementation of the old AP/EAP for the new hardware, or just FSD. It's entirely possible that if you don't have FSD, it runs a NN same as the old.
The current limitations with the old computer are not about response times - it can rapidly detect and respond. It's about making complex decisions about what to do when the instructions aren't simply "stay in lane." It's the advance planning side of things that requires the new computer for FSD, not the ability to veer left when someone cuts you off from the right.
Yes and no. Sure, they will eventually stop updating the software & NN for the old computer. But I don't believe that will be soon - I'm confident it will be years. I don't know if they'll keep updating for decades, but they'll definitely maintain it for years. There's a sizable fleet out there.
And once the software has been updated for many years, it's usually pretty damn stable. For mission-critical stuff, those of us in IT don't choose to run the "latest and greatest." We usually used the most tried-and-true-and-tested versions.
I'd be concerned if, say, 10 years down the line, there were major changes in how roads are marked. This could be done to make FSD easier for a variety of systems. Would Tesla update their software, doing new training, for 12-year-old cars? I don't know. Or let's say you want to take your car to a new country, where driving laws are different. Or hey, if driving laws change 12 years from now - who will update the software for the old computer version to comply with the new laws?
Yes but I do not believe he was trying to take the exit, he was intending to stay in his own lane and so it was Autosteer and TACC that made the error here (not NOA). Autosteer still needs to recognize exits/medians to avoid crashing into them.
I last drove a 2.5 HW Tesla about a month ago, on the latest software, and it still regularly did the thing where it tried to stay in the middle of the lane no matter how wide the “lane” got. There were a few close calls with medians that made me nervous enough to take over. This was without NOA turned on, btw. So pure Autosteer/TACC.
Perhaps to make everyone happy they can let each driver decide if they want to use the new or the old NN on their HW3. In our minds. My car will be on the new NN and your car can be on the old one as per your preference. Everyone wins!
diamond.g
Active Member
I think the lane centering is the strongest and weakest part about Tesla's Autosteer. Strong because it doesn't ping pong or drift (much) like other vendors implementations. Weak because it does lane centering to a fault and exposes the weakness of doing that when lane markings are not clear for exits/merges.Yes but I do not believe he was trying to take the exit, he was intending to stay in his own lane and so it was Autosteer and TACC that made the error here (not NOA). Autosteer still needs to recognize exits/medians to avoid crashing into them.
I last drove a 2.5 HW Tesla about a month ago, on the latest software, and it still regularly did the thing where it tried to stay in the middle of the lane no matter how wide the “lane” got. There were a few close calls with medians that made me nervous enough to take over. This was without NOA turned on, btw. So pure Autosteer/TACC.
Perhaps to make everyone happy they can let each driver decide if they want to use the new or the old NN on their HW3. In our minds. My car will be on the new NN and your car can be on the old one as per your preference. Everyone wins!
It's my understanding that AP2/2.5 only runs with reduced resolution, because that's how the NN is compiled. Plus, in theory, all those other cameras are always active for safety functionality anyways... (at least now that they're in use for them)
Knightshade
Well-Known Member
They run in reduced resolution because the 2.x computer is incapable of processing more than ~200 fps total... and 8 cameras times 30 fps is 240.
This wasn't an issue when they were only going to use 4 cameras for EAP as originally intended, since 4x30 is only 120.
But since they realized they'd need all 8 for things like NoA and enhanced summon to work so they had to go reduced resolution.
None of that matters for the new AP though since it doesn't do any of those things. New AP has no use whatsoever for anything but the front 2 cameras.
If you hit the turn signal it won't change lanes on its own, you need to break out of autosteer and do it yourself- but it can comfortably still handle the 2 side cameras on the side you have the turn signal on toward at full resolution for blind spot detection no problem.
Safety functions still need "all" the cameras, and are always on, which means that probably everything is running at reduced resolution on HW2.x no matter whether you have (E)AP active or not.
Knightshade
Well-Known Member
No, they don't.
AEB for example never needs anything but the front 2 cameras (and the radar).
Blindspot warning as another example never needs anything but the 1 rear-looking side cameras on the side you're turning toward and the 1 rear camera (and the ultrasonics)
Is this true though? My AP only vehicle (even with AP off) is constantly updating the screen with images of what cars are to the side/behind/coming up along side me. Wouldn't the computer be working to continually analyze all the cameras to determine what's there to display the info (even if it's not actually going to initiate a lane change)?
Knightshade
Well-Known Member
The amount of info needed for "is there a car kinda somewhere around here so I can show it on the display" is vastly different from "I need to know precisely where every vehicle is because I'm going to automatically be piloting the car between them" (again remember the "all 8 cameras" thing only got enabled in V9, when Nav on AP became a thing- using only 4 was apparently plenty for all previous EAP features)
Moreover the amount of processing needed is vastly different since new-AP doesn't ever need to care if there's cars in nearby lanes nor ever have to make a single decision based on such info... (while EAP and especially FSD sure do).
blind-spot warning will care- but only about the cars in the lane you're trying to turn into, and it only needs to be paying close attention to 2 cameras for that.
Last edited:
While theoretically possible, I doubt the screen GUI does anything other than leech off the data from what the vision system is seeing
Similar threads
