-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Why ditch the Model 3 with the longest range?
- Thread starter ishabazz
- Start date
... in exactly the same way my existing software stack does not know how to use the extra CPU cores and speed every time I buy a new Mac.
ModelNforNerd
Active Member
I have a theory.
the "new" range on the LR AWD was seen as a direct threat to the S and X, and that may have shown a little in recent S/X sales numbers.
If those cars get their rumored refresh soon, the LR AWD may come back out of hiding, as the S and X will get a range bump to eliminate that overlap.
the "new" range on the LR AWD was seen as a direct threat to the S and X, and that may have shown a little in recent S/X sales numbers.
If those cars get their rumored refresh soon, the LR AWD may come back out of hiding, as the S and X will get a range bump to eliminate that overlap.
Knightshade
Well-Known Member
It's actually nothing like that.
On your mac you're just getting a faster version of the same basic architecture. The software can use the extra speed (it MAYBE can use the extra cores, depending on how multi-core aware it is but that's getting into the weeds).
HW3 is a completely different architecture than HW2.x
So when it's running the 2.x NN it's running in emulation mode. Which is a huge performance hit compared to running native code- and means it's pretending to be the original HW rather than having direct native access to the newer/faster HW.
Which is why Elon mentioned HW3 cars are actually slightly worse than native 2.5 cars right now.
It's a bit like when Mac first switched from PowerPC to Intel chips- even though the new chips were more powerful, the performance jump didn't exist for a lot of old software since it was largely running emulated on the new chip- and only as more things were re-done native to the new HW did the speed jumps really show themselves.
In this case it appears that advanced summon will be coming in the 2.5 code, and since that's the last feature ever promised for EAP, that's likely end of life for the 2.5 codebase, other than small tweaks/optimizations... though new HW3 cars can easily continue to run that in emulation.
Only folks who paid for FSD would need the HW3 codebase (that isn't out yet) and would thus benefit from the HW3 hardware.
Elon has stated multiple times only FSD will need HW3.
It's offered offline, but the reason they are downplaying it is that they are battery constrained and the RWD version uses a lot of batteries, but doesn't add extra profit the way the AWD version does. If they weren't battery constrained, it would be offered.
Knightshade
Well-Known Member
Which part, specifically, of my post do your years at Evil Medical School compel you to doubt?
The part where you assume that self driving is somehow different from Autopilot. The limit between them is purely marketing-driven. The hardware and software behind them are precisely the same, with FSD being nothing more than a functional superset of Autopilot.
This type of automation does essentially two things: it recognizes the surrounding environment (which, in Tesla’s case is mostly image processing) and then runs a decision-making algorithm (i.e. controls the steering, braking and acceleration) using the output of the environment recognition process as inputs. That’s all there is to it; it’s conceptually pretty simple but it’s computationally intensive.
Tesla’s most obvious problem is the fact that the outputs of the environment recognition module are largely *sugar*. Without better environment recognition, the decision-making algorithms can’t do much to improve the overall functionality (I take that back, they could choose to avoid charging into already-identified obstacles for example).
The above-mentioned environment recognition can be improved in one of two ways:
1. Throw computing power at it (which is what Tesla is doing) or
2. Use different/additional sensor suites that improve the quality and quantity of input data, without requiring extra computing power (such as using lidar for depth perception as opposed to determining it from differential image recognition)
Yeah, I don’t work for Tesla and I don’t know exactly what they’re building but, based on their observed functionality and progress I’m very much not impressed by the product in its current state and I believe their incremental improvements will be a lot less spectacular than people think they will be. And I bet that if they ever released their source code, a lot of AI experts would be shaking their heads ...
Edit re #1 above:
Elon’s comment regarding HW2.5 running at 80% capacity vs HW3 running at 5% capacity tells you that they expect some 16x throughput for their current bottleneck, the image recognition algorithm.
That’s what we call in the industry “a little bit better”. In my opinion, they’d need a lot more improvement than that in order to process frames of materially-higher resolution and have a chance of getting somewhat close to “self driving”.
SammichLover
Banned
The problem domain size is a LOT different. With AP you are holding the lane, that's it. When you start deciding to change lanes, that's a massively larger problem. Thus requiring a lot more hardware to run your algorithms involved.
Which outputs? What you the driver (and passengers) see on the screen?
Knightshade
Well-Known Member
The very post to which you were replying is suggesting, based on literally what Tesla/Elon have told us, that there is going to be a HW/SW difference regarding HW3, which was the point to which the post in question was speaking.
HW2.x cars will stay on 2.x if they didn't pay for FSD. And that's fine since the 2.x NN is feature complete once advanced summon is released widely.
Those who bought FSD (and new cars going forward) will come with HW3, since FSD will have a much larger/more advanced NN running to offer more advanced features.
The HW3 cars with no FSD purchase can just keep running the 2.x code in emulation, since it doesn't need/get any new features and it's probably not worth porting the entire thing over to HW3 native for a subset of customers that wouldn't benefit from it anyway.
That is even more the case given EAP isn't sold anymore, so all new customers will either be FSD (get and need HW3 and HW3 native code) or basic AP (which won't even be taxing on HW2 systems).
again, given HW3 is running non-native code, that comparison is pretty worthless compared to what it can do with native code.
I find it hard to imagine that a car with HW3 installed but FSD not purchased will never benefit down the road from software improvements. While FSD is their largest objective, a computer that is that much more powerful creates a lot of opportunity for improvement of other systems in the car, such as emergency braking and forward collision alerts. When V2 came out nobody knew that Tesla would be coming out with Dashcam and Sentry down the road but would exclude V2 hardware from using it. I don't see why we would think that will never happen again with HW3 once it becomes the main hardware platform. Are they really going to hold back development of non FSD related benefits just because there are some V2 and V2.5 cars still running out there?
Knightshade
Well-Known Member
Yeah- see, they don't though....unless the existing 2.x computer is maxxed out when executing those functions.
Which there's 0 evidence actually happens.
According to Elon the current computer is only at 80% capacity doing everything in the current EAP package including NoA.
So doing anything in the much more limited feature set of basic AP, or no AP at all, shouldn't be very taxing to it, and wouldn't benefit at all from "faster computer"
A top of the line brand new PC can run original Pac Man at full speed. So can a 5 year old PC. Or a 10 year old one. The code simply isn't that demanding.
There's no evidence the computer is the issue there though... and since they're making no changes to wiring or sensors with the HW3 change we've no evidence 2.0 cars will magically gain those features with the HW3 upgrade.
On the contrary the evidence suggests it's the ability of 2.0 cameras to do proper color video that is restricting those features (Teslatap suggests a monochrome version may yet appear on 2.0 cars)
HW3 can't fix that.
If you can provide some examples of non-FSD benefits where the current computer is not capable then I'm perfecting willing to accept that might happen.
But I've yet to see anybody come up with one.
Raphymartinez
Member
Ordered a LR RWD this past Friday over the phone and it does come with AP bundled. It is Red with white interior and 19’s for $49,000. AWD version of the same spec is $54,000 right now
SIMD and associative processors are not my specialty, but I can tell you that, in such cases, significant improvement can only come from switching to a different massively parallel SOC architecture. Assuming that HW 2.5 is currently based on generic CPUs/GPUs, it would be possible to see a substantial improvement. However, I don’t see how one could “emulate” a traditional algorithm with “non-native” code on an associative system, which leads me to believe that elon’s 80%-5% comment hints at a less drastic upgrade.
On an unrelated note, I tried the newest Audi driver assistance package today, before I ordered my next car, and the adaptive cruise is way way better than that in Teslas. Super smooth and intuitive in comparison. This is likely a case of doing it the easy way (proper radars) vs. the obsessive Tesla insistence that it be based primarily on computer vision.
SammichLover
Banned
Oh, look, a statement I don’t agree with, let me give this guy a quick insult ...
Did this approach ever work well for you?
SammichLover
Banned
“An absurd, bat guano assertion, probably dreamt up to just be contrary” is a special kind of “opinion” and thus receives a special kind of assessment.
Take it as an insult if you wish, but your statement isn’t you.
sreams
Member
Similar threads
