-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
HW4 & HW5 Discussion
GhostSkater
Member
Worse, afaik he quit his job a while back and FSD testing is all he is trying to doPoor guy just “upgraded” to a Model Y
Or was
I am sure HW4 cars have some form of FSD which approaches the performance levels of HW3 FSD.
But we need someone who owns a HW4 car to confirm this.
My interpretation is there will not be a HW4 specific version of FSD which is better than HW3 FSD for at least 6 months. The likely constraints are data and compute. While Tesla is compute constrained is is better to keep training HW3 versions of FSD.
HW4 has better cameras with more resolution and a higher frame rate, so the issue is the video feed is better than what is needed by HW3.
I would be very surprised to find out that FSD doesn't work at all on HW4. Worst case scenario, HW4 is constrained to an older version of FSD for 6 months,
But we need someone who owns a HW4 car to confirm this.
My interpretation is there will not be a HW4 specific version of FSD which is better than HW3 FSD for at least 6 months. The likely constraints are data and compute. While Tesla is compute constrained is is better to keep training HW3 versions of FSD.
HW4 has better cameras with more resolution and a higher frame rate, so the issue is the video feed is better than what is needed by HW3.
I would be very surprised to find out that FSD doesn't work at all on HW4. Worst case scenario, HW4 is constrained to an older version of FSD for 6 months,
I think worse case, they could do some kind of processing to make the camera's input similar (either up-rez the old cameras or down-rez the new ones to match).
Another thought is, in regards to NN training, with the new cluster of 10,000 H100s coming online, while still having the old 14,000 A100s cluster, means they could use the new, more powerful cluster to train V12 and leave the old cluster to train HW4 for the V11.x or V12.
Another thought is, in regards to NN training, with the new cluster of 10,000 H100s coming online, while still having the old 14,000 A100s cluster, means they could use the new, more powerful cluster to train V12 and leave the old cluster to train HW4 for the V11.x or V12.
Yes, I also thought that.
I am most interested to know if anyone with a HW4 car has FSD and how that is going, I find it very hard to believe HW4 cars don't have FSD.
What I think would work is a down-rez of the HW4 video feed to a standard compatible with NNs trained on HW3. Short term that only means one set of NNs being trained. HW4 data is simply banked for future use.
I might be wrong, but I don't understand why I am wrong, perhaps the down-rez is harder than we think it is, but all of us have done it with photos.
Last edited:
The down res might not work because they are feeding the neural net the raw photon count, not a processed image.
Yeeees, but you’d think the very smart Tesla engineers could figure it out. Ah well, we’re just going to have to wait to get confirm somewhere of people with HW4 and FSD beta be it v11 or v12
powertoold
Active Member
True, but they can simply drop every other raw pixel value (for example), no actual "processing" needed.
Artful Dodger
"Neko no me"
Six months is probaly the minium amount of time it will take to collect data from the HW4 fleet to create a viable implementation of FSD. If this approach works, in the future we could expect similar time frames for other models. Getting to volume production for CT will be important, and will likely go even faster with Model 2/Gen 3.
Cheers!
Artful Dodger
"Neko no me"
If the market wakes up and realizes the true potential of FSD ($$$$) and $TSLA takes off to the moon, you can buy a HW3 equipped car for fun to use FSD now.
Kinda turns the used-car market on it's head for previously-owned Teslas too, don't it?
Imagine if the Robotaxi Fleet is HDW3-only to begin? Happy now that you bought a Tesla in 2016.10-23.07?
Cheers to the the Unexpected!
dhanson865
Well-Known Member
No, the point is give the nueral nets unprocessed video (raw photons) so that they don't have any training towards a system that isn't needed and latency is as low as possible.
It's not that it doesn't see video, it's that the video it sees didn't go somewhere else between the camera and the neural net that is going to make the final decision.
Get rid of uneeded layers and make it as simple a flow chart as possible.
Andy O
Member
Since the Neural Network is trained by watching videos of good human driving, how will Tesla keep it current when there are no human drivers left, only FSD cars?
Seriously...that's actually quite interesting. Right now we are in the early stages of transition...mostly human drivers with a few robots, so the robots need to mimic the best humans. Over time, if eventually 100% of cars on the road are self driving, then driving patterns and rules should change. 100% robot drivers acting like perfect human drivers is good (and better than today), but not optimized. Robotaxis with surround cameras and high speed computing will have better awareness and reaction times, and be capable of much more: higher speeds, closer spacing, tight merging, passing through narrow gaps in cross traffic, etc. But how can we flip the switch to get there unless all the cars (at least in a given area) upgrade at the same time and/or know and trust each other's capabilities...
Related to Tesla and TSLA specifically, I think this means that for a truly great future, 100% of those future 100% robot cars need to be running Tesla's FSD.
....Bullish!
Some great commentary here about the V12 stream. I wish some of the authorities (posters here) on domains relative to autonomous driving would speak up a little bit more, though. Maybe I can summon you by posting wrong information, because I'm someone who is wrong and I am posting on the internet!
- V12 will reach its own maximum on performance on HW3 and a different maximum on HW4. There will be V13, V14, V15, ...
- V12, and the whole system behind it, raises the maximum level of the system to be at least that of some weighted average of the best training dataset they can gather from the fleet, which will come from the best drivers they can collect data from (they can oversample though)
- V12 may exceed the performance of those drivers simply because it never gets tired, possibly enabling superhuman driving just by negating human limitations of the best driver on the road
- V12 may need another layer on top of it (maybe we're in V13 now) to become superhuman
- V12 may reduce crash risk far beyond what Tesla reports for basic autopilot even if it's not at the level of Waymo or Cruise and still classified as level 2
- The people who said the cores were maxed out need to reevaluate their lives
I have HW4 and FSD is not enabled. It feels like EAP before FSD was introduced so hard to judge it by anything.
I think people should understand what diffusion is before speculating on why HW4 video feed can or cannot be down rez.
Diffusion is taking pixels and working backwards to form an image, so it's more of a predictive model. So basically a good diffusion model pretty much eliminates the need for high resolution. In fact the lowest number of pixels that can predict the correct image is best due to less compute needed. So if Tesla can achieve extremely good predictive accuracy using 1.3mp, then I suspect they will need to matrix the high resolution up and convert it down to 1.3mp to free up computer power.
I should have checked TMC before posting the question.
This post has the most interesting information:- Do any HW4 Vehicles have FSD Beta activated ?
This post has the most interesting information:- Do any HW4 Vehicles have FSD Beta activated ?
Also this one:- Do any HW4 Vehicles have FSD Beta activated ?
Last edited:
No, I don't think they need to learn that at all before speculating.
Yea, diffusion models denoise to get something a human can look at and understand.
Yea, if you could put one gigantic block across 1.3mp in a single RGB value and denoise it, that would take less compute.
Putting my neck out here trying to get some of you in the shadows to come out and comment! My company runs a real world model on extremely real world AI using training data from cameras at different resolutions, lenses, with different ISPs, etc. We have a human in the loop to make any decisions. Works great! Far better than any human, and we know that, because we used to pay them to do the job and our insurance costs are projected to decrease!
Threads of the day:
FSD discussion Anyone buying a CT or Highland now will be in the FSD wilderness until Spring 24
Tesla Network prerequisites For those thinking beyond FSD or here or Super bull thread
Industrial chic:
Edit - source
FSD discussion Anyone buying a CT or Highland now will be in the FSD wilderness until Spring 24
Tesla Network prerequisites For those thinking beyond FSD or here or Super bull thread
Industrial chic:
Edit - source
Last edited:
ToTheStars
Member
I think the main point here is that the neural network will automagically discover any future improvements they could have found. Thus it will be hard to beat. That's the nature of neural networks, they will discover what they need to discover, as long as the network is large enough and you have enough data and you train them long enough. And Tesla has plenty of all of these.
Similar threads
