Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Hardware 3.0 Increase Non-EAP/FSD Features?

This site may earn commission on affiliate links.
Does the hardware system (2, 2.5 or 3) have anything to do with the car systems, outside of Enhanced Autopilot or Full Self-Driving? I've placed an order for a Model 3 Dual Motor Performance with EAP and the FSD option. I'm fully aware the FSD will likely not come to fruition during my ownership of the car, but I was originally getting it for incremental updates to the EAP which could require the newer hardware 3.0. I'm now questioning how much I would really use EAP. Following this progression, if I'm not going to use EAP, and don't expect FSD to materialize soon, the only reason to get this would be for the hardware 3.0 chip. If this upgraded hardware doesn't have any effect at all on other systerm (navigation, streaming content, etc), then I should probably cancel the EAP/FSD. If that hardware does increase the capabilities of the systems reflected in the 15" screen, it becomes a ROI question.

So, in a nutshell - do the hardware system upgrades benefit the goodies on the console screen?
 
It's extremely unlikely that HW3 is related to the center console screen. The infotainment computer (CID/MCU) is a separate chip, and nothing shown in the latest firmware leaks suggests any different.

But with that said, it's very much possible (and probably inevitable) that HW3.0 "EAP" may perform better than HW2.0/2.5, even though Tesla considers them equivalent. For example, right now, HW2.5 and HW2.0 radars have different ranges and reaction to overpass braking, even though Tesla considers them functionally equivalent. And MCU2 vs MCU1 is considered equivalent in Tesla's mind, but owners of MCU1 beg to differ.

I highly doubt Tesla would go out of their way to cripple EAP non-FSD HW3.0. Even if they run the exact same neural nets, it's already been documented that HW2.0 is kind of quirky when it comes to things like the rear camera reliability. Are there other workarounds on the Nvidia platform that we don't really know about, but are holding HW2.x back? Maybe.

With that said though, the retrofit schedule might not be something that you like. Surely all available HW3.0 supply will be going into new production cars. Tesla is not going to compromise their production numbers in order to give existing customer retrofits!

Long story short, if you can afford it, it's probably reasonable right now to pre-purchase FSD to make sure you're on the bleeding edge. It seems inevitable to me that HW3 will be more capable than HW2 in all aspects, whether or not it's FSD feature related.
 
I'll agree with @chillaban, what you see is coming from the display computer, not the navigation computer.
I've not heard of any rumors that there is anything else that the upgrade will impact.

As far as in the "time you own the car" the software that uses the new computer is out being tested right now.
 
Does the hardware system (2, 2.5 or 3) have anything to do with the car systems, outside of Enhanced Autopilot or Full Self-Driving? I've placed an order for a Model 3 Dual Motor Performance with EAP and the FSD option. I'm fully aware the FSD will likely not come to fruition during my ownership of the car, but I was originally getting it for incremental updates to the EAP which could require the newer hardware 3.0. I'm now questioning how much I would really use EAP. Following this progression, if I'm not going to use EAP, and don't expect FSD to materialize soon, the only reason to get this would be for the hardware 3.0 chip. If this upgraded hardware doesn't have any effect at all on other systerm (navigation, streaming content, etc), then I should probably cancel the EAP/FSD. If that hardware does increase the capabilities of the systems reflected in the 15" screen, it becomes a ROI question.

So, in a nutshell - do the hardware system upgrades benefit the goodies on the console screen?
You’ll want EAP. It’s great...not perfect yet, but great nonetheless.
 
What Tesla has said is the upgrade is only for FSD cars, it's not needed on EAP cars.

The most direct reading of that is that EAP won't "gain" anything from HW3, and HW3 isn't needed to deliver all the features promised to EAP buyers (hence no free upgrade for them, but free upgrade for FSD buyers)

The simplest way for Tesla to make that true is simply label any new features enabled by HW3 as FSD only features (and unlocking them only for HW3/FSD-paid owners) - and indeed Tesla has said the first FSD-Only features would arrive about the same time as HW3 does.


Now, some folks think that EAP might still work "better" on HW3 than 2.5 even if it doesn't get any new features... that's certainly possible from a technical perspective but I'm not sure how many NN branches they want to maintain on different HW (versus running the same one on all 2.x and 3.x HW for non-FSD owners).
 
If HW3 can process frames 10x faster, I can't see them intentionally neutering it while using an EAP feature. That would potentially create a safety issue. If there's something HW3 can tag more precisely, they should use it to the best of its ability, no?

For that reason, I do expect non-FSD features that use computer vision to benefit in some way. Even Advanced Summon or Automatic Parking might benefit. If my experience with AP1 is any guide, they certainly do seem to direct the majority of their development time to the latest hardware.

We'll never have a perfect system. It's really just the number of 9's will have trailing in 99.9999% accurate. I believe HW3 will ultimately be "safer" than HW2.5 in all areas. Incrementally (and maybe unnoticeable) in most cases, but enough to potentially prevent an accident in a complex 1 in 100,000 situation.

I don't think it will affect the center display unless it might be able to more accurately depict cars and pedestrians. With sufficient NN data, it could (for example) differentiate between the makes of cars and more accurately display them. So it might be better for situational awareness. But that's more about quality of data than the GPU powering infotainment. It might be better at handling automatic wipers or high beams. And I can see Sentry mode improving with recognizing threats or the owner more accurately.
 
  • Like
Reactions: UngaBunga1989
If HW3 can process frames 10x faster, I can't see them intentionally neutering it while using an EAP feature. That would potentially create a safety issue. If there's something HW3 can tag more precisely, they should use it to the best of its ability, no?

No...because then they'd need to run a whole additional code base when they already clearly don't have enough programmers.

Right now they have the NN that runs on AP2.x stuff doing EAP functions. EAP is feature complete (other than turning off requiring you to confirm lane changes, and maybe some summon improvement).

No doubt they'll keep improving that NN over time.

HW3 can run the same 2.x NN on EAP only cars quite easily. Only need to maintain one codebase for EAP cars. And it supports Elons statement HW3 isn't needed for anything to do with EAP (and thus no reason to wait for HW3 if you only want EAP)


Then there'll be a new codebase/NN that is for FSD cars intended to run on HW3 that is a lot more capable and will offer additional features that EAP does not.

If they wanted to have a third NN (EAP that only works on HW3, but doesn't do any of the FSD stuff) that's potentially adding a lot of complexity.


By making all features beyond what HW2.x EAP does designated as FSD features they finally differentiate FSD, avoid needing to create and maintain a third NN codebase, and they avoid any obligation to provide free HW3 upgrades to EAP 2.x buyers.
 
Long term HW3 will definitely be better also for non-FSD functionality. Tesla will most likely never admit it though unless they decide to upgrade non-FSD cars too.

The reason is software. In the early days you can run EAP like before on the same software code and neural networks, just using different device drivers at lower levels for it to work on both platforms. But at some point in FSD development there will be enhancements required for FSD that are shared with EAP/Security that will not be backward compatible with HW2.x.

After that point, improving HW2.x would be twice the work for Tesla customers that don't pay for it and does not generate sales. Not very motivating. Then they just split the old and new AP code and stop updating that part for older cars. Just like AP1 no longer gets any new features or significant improvements.
 
Tesla will most likely never admit it though unless they decide to upgrade non-FSD cars too.
Maybe. Maybe not. I think there is a real possibility that Tesla will say AP2.5 checks all the EAP functionality boxes and does so with an accident rate that is 2.5x better than average human drivers. If you want to pay for the AP3.0 hardware upgrade, which is not necessary, we can check all the boxes with an accident rate that is 15x better than average human drivers....
 
Maybe. Maybe not. I think there is a real possibility that Tesla will say AP2.5 checks all the EAP functionality boxes and does so with an accident rate that is 2.5x better than average human drivers. If you want to pay for the AP3.0 hardware upgrade, which is not necessary, we can check all the boxes with an accident rate that is 15x better than average human drivers....

Then they better remove the beta designation.
 
  • Like
Reactions: pilotSteve
No...because then they'd need to run a whole additional code base when they already clearly don't have enough programmers.

Right now they have the NN that runs on AP2.x stuff doing EAP functions. EAP is feature complete (other than turning off requiring you to confirm lane changes, and maybe some summon improvement).

No doubt they'll keep improving that NN over time.

HW3 can run the same 2.x NN on EAP only cars quite easily. Only need to maintain one codebase for EAP cars. And it supports Elons statement HW3 isn't needed for anything to do with EAP (and thus no reason to wait for HW3 if you only want EAP)


Then there'll be a new codebase/NN that is for FSD cars intended to run on HW3 that is a lot more capable and will offer additional features that EAP does not.

If they wanted to have a third NN (EAP that only works on HW3, but doesn't do any of the FSD stuff) that's potentially adding a lot of complexity.


By making all features beyond what HW2.x EAP does designated as FSD features they finally differentiate FSD, avoid needing to create and maintain a third NN codebase, and they avoid any obligation to provide free HW3 upgrades to EAP 2.x buyers.

I don’t have inside information nor decompile the code, but Karpathy did say that they can control the running time of neural nets just by removing layers. It is very straightforward and allows them to maintain the same code base. That might be the strategy going forward at some point... for the various NN’s, cut down the number of layers to “fit” AP2.x processing power and get a correspondingly poorer answer, but good enough for EAP.
 
I don’t have inside information nor decompile the code, but Karpathy did say that they can control the running time of neural nets just by removing layers. It is very straightforward and allows them to maintain the same code base. That might be the strategy going forward at some point... for the various NN’s, cut down the number of layers to “fit” AP2.x processing power and get a correspondingly poorer answer, but good enough for EAP.

Do you have a quote on that?

Because from the one I've seen it's not that HW3 just lets them run current codebase NNs faster, but it lets them run different, much larger, NNs-


Andrew Karpathy said:
This upgrade allows us to not just run the current neural networks faster, but more importantly, it would allow us to deploy much larger computational and more expansive networks to the fleet. The reason this is important is that it is a common finding in the industry that as you make the networks bigger, the accuracy of their prediction increases with the added capacity. Now we’re currently at a place where we’ve trained large networks that work very well, but are not able to deploy them to the fleet due to computational constraints. So all of these will change with the next iteration of the hardware, and it’s a massive step improvement in computing capability, and the team is incredibly excited to get these networks out there


 
Outside EAP/FSD is the safety features like emergency braking and side collision avoidance. HW3 could easily improve on those. I've been thinking if HW3 gives you EAP using the larger neural net it might be much more sure-footed than the current version. Our X has FSD prepurchased, while the 3 has only EAP. Maybe eventually we'll see if there's a big difference in EAP performance.
 
I fully expect HW3 to increase/improve non-EAP and FSD features.

Some examples:

Dash Cam Capability -> Currently HW2 can't do the Dash Cam.

Sign recognition -> I haven't seen anything from Tesla saying HW2/HW2.5 will do sign recognition, but I have seen it claimed that HW3 can do sign recognition.

Someone else brought up a good point about Tesla not wanting to support two code bases, but HW3 is so different that in reality they do have two code bases for quite a bit of code that's on the ADAS computer.

I also don't see any issues with having a more complicated neural net used while driving with a HW3 computer even in purely manual driving. If you treat it as a module where the inputs and outputs were similar enough. You'd simply have more outputs from the more advanced HW3 network (sign detection, etc).

The way I tend to think of HW3 is it's for Safety+EAP+FSD where there is a single code base, and it's only SW flags that allow for any EAP or FSD type feature.

Where HW2/HW2.5 is strictly limited to Safety+EAP, and runs a less complicated code base (in terms of the driving computer). So the neural nets aren't as advanced.

I expect that HW2/HW2.5 will mostly enter a maintenance only mode within 6 months. I expect that most of the bugs that currently exist (like when the rear camera doesn't work) won't ever be fixed.

Tesla will concentrate their efforts to incentive people to upgrade to HW3, and part of that will be having features to entice people to upgrade. Tesla will want to get as many HW2/HW2.5 people over as possible.
 
Last edited:
I fully expect HW3 to increase/improve non-EAP and FSD features.

Some examples:

Dash Cam Capability -> Currently HW2 can't do the Dash Cam.

Which never made any sense.

Even less sense when you consider Elon said HW2 will get Sentry mode using all the cameras.


Sign recognition -> I haven't seen anything from Tesla saying HW2/HW2.5 will do sign recognition, but I have seen it claimed that HW3 can do sign recognition.

...as part of FSD.

EAP was never intended to read stop signs and such- FSD is.

Someone else brought up a good point about Tesla not wanting to support two code bases, but HW3 is so different that in reality they do have two code bases for quite a bit of code that's on the ADAS computer.

The point was avoiding three codebases.

EAP NN on HW2.x
EAP NN on HW3
FSD NN on HW3

That middle one is not needed if they just run the 2.x EAP NN on the 3 HW (which Karpathy already said works fine) and only give any advanced/improved features to FSD owners who will all be on HW3.


Tesla will concentrate their efforts to incentive people to upgrade to HW3, and part of that will be having features to entice people to upgrade. Tesla will want to get as many HW2/HW2.5 people over as possible.


The best way to do that would be to offer FSD-only features people want- since that's the only path to getting HW3 in a 2.x car we're currently aware of.

It also avoids people who paid for EAP saying they deserve free upgrades too.
 
EAP was never intended to read stop signs and such- FSD is.

But, we can agree that EAP should recognize speed limit signs right? Or at the very least it's database of speed limit signs should be significantly improved. In my experience the navigation system has a really bad database of speed limits, and it's constantly wrong.

We're pretty far along in the development of HW2/HW2.5, and they still don't have speed limit sign recognition. But, it's been claimed that Tesla has working sign recognition with NN's running on HW3.

So it seems doubtful to me that we'll see any kind of speed sign recognition with HW 2.X, and instead Tesla will try to get sign recognition within HW3 as quickly as possible. That way it can be released to fleet of HW3 vehicles (once they start making them) where it can also be used to improve Tesla's database of speed limit signs to improve the situation for people with HW 2.X.

EAP and FSD being separate code bases doesn't support the approach Tesla is taking to FSD. Despite how they sold them they're taking an incremental approach.

In my own experience with NN's it's pretty painless to have separate ones depending on the needs of it, and the capability of the hardware that's running the neural network. It's not like having separate code bases for an application.

I see no reason at all to have a separate codebase for FSD. Instead it's simply going to be growing the EAP code base, and HW 2.x cars will just get left behind. Maybe a neutered version of it.

So I see

Simplified NN on HW2.X that will largely be in maintenance mode.
Advanced NN on HW3 that will be under active development

The disappearance of any kind of separation between EAP/FSD from within the codebase on a HW3 vehicle. Now I'm not saying that Tesla won't monetize some kind of FSD feature like limited L3 driving on the freeway. But, it's only going to be a SW enable flag.

It's also going to disappear because FSD has always been a false promise. This pie in the sky thing that the car is simply not capable of doing. So it makes way more sense to me to abandon that approach entirely, and to leverage the fleet of vehicles to improve the entire Autopilot frame work gradually.

Go to a single codebase with HW3
Improve NoA (an EAP feature) to where it's basically L3 capable, but only allow L3 to people who pay specifically for it (like FSD owners today). Try to get it within 2019 so we're not beat by other car companies.
Add rear radar with the Model Y (to augment the cameras)
Add more and more things to the database of stuff the car can recognize.
 
  • Informative
Reactions: pilotSteve