HW3 soon moving from emulation to native mode

[Kmartyn]

The suggestion of this article, is that the software rewrite mentioned by Elon in yesterday’s Third row video, is a move from emulation, to native mode on HW3.

Tesla Autopilot Mystery Solved — HW3 Full Potential Soon To Be Unlocked | CleanTechnica

This rewrite also combines all the cameras from individual analysis of still shots to one real-time 3D video, with an order of magnitude improvement in labeling accuracy.

This looks to be the step that will give us feature complete FSD, by the end of 2020.

Really hope Tesla can find a way to pick up the pace of HW3 upgrades.

 

[linux-works]

ah, I didn't catch this part! now I'm getting a better picture of what's going on.

so, the 're-write' is primarily a port of the current code - but removing the EMULATION LAYER and making the api calls direct instead of indirect and abstracted.

this would improve speed, remove any middlware translations (some api style are not 1:1 so adapters in software are needed and those waste time and memory) and let them tune it to the actual target hardware.

lots of companies would just ship the emulated code and leave it at that! not kidding (sigh). good that tesla is biting the bullet and now converting their emulation layer to native.

at least that's what I got from that article; if we can trust its accuracy.

bad news is that they seem to have gone with ONLY hw3 and so the new code will not run, eventually (at all) on non-hw3 platforms.

that's probably the right solution, otherwise you keep bringing in so many IFDEFs (lol) that the code becomes hard to maintain and support.

people without hw3 are in bad shape; lets hope tesla does the right thing for those owners.

 

[Kmartyn]

ah, I didn't catch this part! now I'm getting a better picture of what's going on.

so, the 're-write' is primarily a port of the current code - but removing the EMULATION LAYER and making the api calls direct instead of indirect and abstracted.

this would improve speed, remove any middlware translations (some api style are not 1:1 so adapters in software are needed and those waste time and memory) and let them tune it to the actual target hardware.

lots of companies would just ship the emulated code and leave it at that! not kidding (sigh). good that tesla is biting the bullet and now converting their emulation layer to native.

at least that's what I got from that article; if we can trust its accuracy.

bad news is that they seem to have gone with ONLY hw3 and so the new code will not run, eventually (at all) on non-hw3 platforms.

that's probably the right solution, otherwise you keep bringing in so many IFDEFs (lol) that the code becomes hard to maintain and support.

people without hw3 are in bad shape; lets hope tesla does the right thing for those owners.

I don’t think it is a port, but rather a native rewrite, merging different functions and putting much more load on the NN.

They are hiring and expanding the team in the following areas:

• Hardware
• Neural Networks
• Autonomy Algorithms
• Code Foundations
• Evaluation Infrastructure

Tesla releases new video of what Autopilot can see, Cybertruck simulation, and more - Electrek
It certainly does sound like all HW3 going forward for FSD, but based on, past Tesla statements, I don’t think that is a surprise.

 

[diplomat33]

Here is part 2 of the podcast where Elon talks about the rewrite and more about AP:

 

[lowtek]

bad news is that they seem to have gone with ONLY hw3 and so the new code will not run, eventually (at all) on non-hw3 platforms.

The best progress is made when you free yourself from the shackles of backwards compatibility.

 

[linux-works]

when I say 'port' I don't just mean 'make it compile' but make semantic changes so that your functionality makes different api calls into the new system rather than calling the old routine and having that routine be a 'stub' that just manages data and sub-calls to translate to the new system.

like going from a VM to native bare metal. that's how I'm reading this, but even that might not be the best analogy. close, though.

 

[diplomat33]

In the podcast I posted above, Elon says that dojo will go online towards the end of this year or early next year. So it sounds like Tesla is focused on doing the rewrite now and then once that is done, they will use dojo to further accelerate learning with the new rewrite.

 

[azred]

The best progress is made when you free yourself from the shackles of backwards compatibility.

Tesla has been selling FSD for over three years with zero product delivered to the early customers. Every feature the 2.0 crowd has to date is part of the old EAP. Providing any FSD features to those customers on 2.0 who have been waiting three years + for any features is hardly backwards compatibility in the sense that I understand the phrase. But perhaps I am misinterpreting your comment.

 

[lowtek]

But perhaps I am misinterpreting your comment.

No, you interpreted it correctly, dropping support for legacy hardware would certainly speed things along.

Tesla has already calculated this, hence the upgrades to HW3 for some.

I'm not sure Tesla has a legal responsibility to the previous generations .. haven't seen the Terms those purchasers agreed to .. unfortunately, the marketing propaganda doesn't count ..

 

[azred]

No, you interpreted it correctly, dropping support for legacy hardware would certainly speed things along.

Tesla has already calculated this, hence the upgrades to HW3 for some.

I'm not sure Tesla has a legal responsibility to the previous generations .. haven't seen the Terms those purchasers agreed to .. unfortunately, the marketing propaganda doesn't count ..

Those of us on 2.0. who purchased EAP and FSD have received nothing for FSD and you aren’t sure Tesla would be liable if they never do? Of course at this rate anyone leasing will be out of their cars and others who flip often will also be out of their cars so the cost is already rapidly declining.

 

[lowtek]

Those of us on 2.0. who purchased EAP and FSD have received nothing for FSD and you aren’t sure Tesla would be liable if they never do? Of course at this rate anyone leasing will be out of their cars and others who flip often will also be out of their cars so the cost is already rapidly declining.

Nope, not sure. Why would I be? I never purchased a HW2 vehicle so therefore wouldn't know what was signed during the purchase.

But I'm quite sure that marketing propaganda won't play much of a part, so don't hang your hat on what you saw on their website or discussed with an employee. It's that purchase agreement that counts.

If Tesla comes out with say the perfect HW3 FSD and you feel like it's owed to you, first, disconnect that feeling as it doesn't matter how you feel about it, only the facts you can prove matter. Second, contact a lawyer and have them review your purchasing agreement. Maybe there would be a favorable settlement for you, that is, if Tesla doesn't bury you in litigation and you go bankrupt first. Maybe that will be the start of a class action lawsuit, maybe not. It could be that maybe, just maybe, you're going to be left behind.

Tesla knows all this, they aren't idiots, the know exactly what they've done, promised, and what their going to do. They likely have great lawyers being a sometimes controversial auto manufacturer doing bleeding edge things. They've already calculated exactly how much moving on to HW3 only, and leaving HW2 behind without the same level of FSD or any FSD at all, is going to cost them in litigation with unhappy HW2 owners. They may find it cheaper to just settle with those HW2 owners who can afford to lawyer up, which would include a non-disclosure agreement effectively putting a lid on it. Kind of like a recall, which is decided upon based on a spreadsheet and not much else. Sadly, we are statistics to be reconciled by the machine.

 

[diplomat33]

Does anybody have any thoughts on how close Tesla might get to say L3 autonomy on the highway with this new AP/FSD rewrite?

 

[boonedocks]

Does anybody have any thoughts on how close Tesla might get to say L3 autonomy on the highway with this new AP/FSD rewrite?

To be honest I would be willing to settle for this on both of my current cars if that’s all I ever get for $$$$ pre paid for FSD. I truly hope Tesla doesn’t wait on City driving whether NOA or not for L3 highway driving to be in production!!!

 

[S4WRXTTCS]

Does anybody have any thoughts on how close Tesla might get to say L3 autonomy on the highway with this new AP/FSD rewrite?

It doesn't change the fact that Tesla has no sensor redundancy for the rear camera or the two back facing cameras on the sides.

 

[Eno Deb]

Uh, has anybody anywhere with inside knowledge said anything about emulation? That seems highly unlikely to me, given the specialized nature of the neural net accelerator. More likely they used some kind of cross-compiler to recompile their existing models for HW3.

 

[diplomat33]

It doesn't change the fact that Tesla has no sensor redundancy for the rear camera or the two back facing cameras on the sides.

True that could be a problem. As I have stated before, the lack of redundancy could mean that Tesla achieves autonomous driving with driver supervision but is not be able to achieve the safety necessary to remove the driver.

 

[Daniel in SD]

It doesn't change the fact that Tesla has no sensor redundancy for the rear camera or the two back facing cameras on the sides.

Why would sensor redundancy be required for L3? If the side or back cameras fail just have the driver take over. Heck, even if the front cameras and RADAR fail you could probably safely have the driver take over 99.9% of the time.

 

[ScubaMal]

I'm to wonder if this might be the reason of the weaving between lanes or (ping ponging) while on Autopilot since my upgrade to HW3. Emulation isn't reacting fast enough to keep the car centered in the lane. Just thinking out loud.

 

[diplomat33]

CleanTechnica feels that the rewrite is a good thing but will set FSD back a year:

"My take: Unlike the conclusion drawn in this article, this is a major delay to the Full Self Driving program. This means the demo we saw at Autonomy Day last year was judged to be not good enough to put into production and they had to go back to the drawing board. As we learned from listening to Andrej Karpathy talk about the tradeoffs between combining things and keeping them independent, putting 3 functions together will make if more efficient, but that will have 2 other effects. The interfaces to all the other parts of the system need to at a minimum be retested and maybe redesigned. Also, the ability to just quickly hand code around a small known issue in planning, perception, and image recognition is removed. Instead, you have to train the neural network around each known issue. I think this is a good change, but it probably sets the system back close to a year."
Timestamped Guide To Part 2 Of Elon Musk Interview By Third Row Tesla | CleanTechnica

 

[linux-works]

I'm to wonder if this might be the reason of the weaving between lanes or (ping ponging) while on Autopilot since my upgrade to HW3. Emulation isn't reacting fast enough to keep the car centered in the lane. Just thinking out loud.

I doubt that's the problem.

keeping a feedback loop ('servo') working - in physical 'car time' is not too hard. its millisconds, not microseconds. easy peasy.