I read Elon’s tweet as “we are still working on a native HW3 codebase. Until that is ready, we have to run some middleware on the HW3 platform which makes the Tesla NN chip look like a GPU, so that the otherwise-unmodified 2.5 code can run on it.” Or something like that.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla autopilot HW3
- Thread starter verygreen
- Start date
I read Elon’s tweet as “we are still working on a native HW3 codebase. Until that is ready, we have to run some middleware on the HW3 platform which makes the Tesla NN chip look like a GPU, so that the otherwise-unmodified 2.5 code can run on it.” Or something like that.
I think it's important to point out what the difference is.
The HW2.5 computer has two SOC's. You can consider a SOC (system on a chip) as taking care of all the house keeping. Things like bringing in all the 9 camera feeds, and talking to other hardware. It has two of them for redundancy, but somehow the second one ended up running Dashcam functionality along with Sentry mode.
Also included within HW2/HW2.5 is a neural network accelerator. This is what does all the heavy lifting in running the deep neural network.
The HW3 computer changed out the SOC, but that wasn't too big of a deal. It simply presented work to do in order to create new drivers for sensors, etc. The hassle of dealing with a different ISP (image signal processor), etc. Nothing that should really impact how well FSD works assuming they don't have any serious bugs. As far as I know everything that worked before still does with exception to dashcam and Sentry Mode.
What really matters is the neural network accelerator as this is what Tesla spend a bunch of time, and resources on creating.
One of the really fun things about Neural networks is they're fairly easy to migrate from one system to another. It's really optimizing them that's a pain, but Tesla took care of that by optimizing the hardware for their type of neural networks.
It's extremely doubtful that the HW3 has any issues running the existing Neural network, and in fact it's probably barely even scratching the surface of what it can do.
The time being spent right now is probably mostly bug fixing, and continuing to bring things to HW3 that may have gotten broken (like Dashcam, sentry mode, and whatever else might be).
There is also likely additional stuff they want to fix/tweak for HW3.
It's very likely they have a New neural network almost ready to go for HW3.
They're so close to parity that there is no reason to rush anything.
As to the pricing structure? It was necessary to allow the features in question to grow.
They were features that really shouldn't have been under EAP in the first place.
The HW2/HW2.5 computer simply doesn't have the processing power needed to do a great job of Enhanced Summons or NoA. It's why both of those features are so late, and that neither one is all that impressive.
My prediction is the EAP people who didn't get FSD are going to be pissed. Luckily Tesla got quite a few of them over to FSD to lessen the blowback.
The first FSD feature of detecting stop lights also doesn't perform all that well on HW2/HW2.5, At least in what we've seen so far.
no. this is incorrect.
There's a "ODM" fuse. once it's blown, no other fuses could be blown after that at all no matter their state. I think fuses can be changed arbitrary before the ODM one is blown though.
I’m thinking concern (or even doubt) over FSD on AP2 / Model 3 retrofits may well be @verygreen ’s one big miscalculation. (Assuming Tesla does retrofits for Model S/X AP 2.5. If they retrofit nobody that would be different.)
While I see the point that Musk could be wrong or lie on Twitter, I just don’t see verygreen’s rationale for Tesla not doing the retrofits on AP2. What? Just because they’d need an adapter cable for a couple of Fakras and a bit in software saying AP2/2.5 and taking a few things into consideration? That’s nothing. That’s no reason to expose the company to significant legal action over that.
Model 3 retrofit no doubt is harder but I think for that Tesla will simply go for the ”replace the whole assembly” route and recycle those computers back to base for repurposing, if it is too hard to do on-site. Again, this too they can not afford not to do.
I do agree at some point Tesla may well strip CPOs/used cars sold through Tesla (as is rumoured) from FSD upgradeability that are more costly to do but that is a completely different story of course, if communicated before purchase.
While I see the point that Musk could be wrong or lie on Twitter, I just don’t see verygreen’s rationale for Tesla not doing the retrofits on AP2. What? Just because they’d need an adapter cable for a couple of Fakras and a bit in software saying AP2/2.5 and taking a few things into consideration? That’s nothing. That’s no reason to expose the company to significant legal action over that.
Model 3 retrofit no doubt is harder but I think for that Tesla will simply go for the ”replace the whole assembly” route and recycle those computers back to base for repurposing, if it is too hard to do on-site. Again, this too they can not afford not to do.
I do agree at some point Tesla may well strip CPOs/used cars sold through Tesla (as is rumoured) from FSD upgradeability that are more costly to do but that is a completely different story of course, if communicated before purchase.
Last edited:
https://electrek.co/2019/03/29/tesla-full-self-driving-computer-retrofit-elon-musk
Retrofits will start when our software is able to take meaningful advantage of the Tesla FSD computer, which is an order of magnitude more capable. For now, it’s slightly disadvantageous to have Tesla FSD computer, as our software is more refined for HW2.
Retrofits will start when our software is able to take meaningful advantage of the Tesla FSD computer, which is an order of magnitude more capable. For now, it’s slightly disadvantageous to have Tesla FSD computer, as our software is more refined for HW2.
Yeah, I covered that aspect on the next line.
"Now they likely locked it down so that the private key (in the SOC chip) can't be changed (as in setting a bunch of the remaining zeros a 1)"
The ODM is the master lock
You confirmed that they locked it.
Hopefully there is some option that neither you or I are seeing because trashing/recycling 100,000 Nvidia PX2 computers is a terrible waste.
Two thoughts I have not yet been able to validate:
1. Tesla may have made the key based on a unique seed value for each unit so that there is not a single key for all. That eliminates the entire fleet being compromised if one key gets out. (Fleet requires the transformation key/ algorithm).
2. Many micros I've worked on have a full reset/ erase function that clears all non volatile storage and then resets the security bits. However, others are OTP only. Not sure which this system is.
1. Tesla may have made the key based on a unique seed value for each unit so that there is not a single key for all. That eliminates the entire fleet being compromised if one key gets out. (Fleet requires the transformation key/ algorithm).
2. Many micros I've worked on have a full reset/ erase function that clears all non volatile storage and then resets the security bits. However, others are OTP only. Not sure which this system is.
It’s wasteful, but in the long run it will save a lot of dollars. Doubt they will have much value to Tesla right now as their new hardware has a minimal marginal costs, consumes less energy and are worse at gathering data. Don’t think it would be a good strategy to use APH4 for FSD and reuse old APH3 for new non-FSD cars.
Why wouldn't Tesla just use old HW2.5 boards as replacement parts? It's not like everyone will upgrade to HW3. I suspect the upgraded parts will be used as spares.
It is also interesting reading this thread because I have never had a car before where people were so upset over whether they would get upgraded to the latest hardware. I understand people paid for it. I did too. I'm just saying I never had this as an "issue" with any other car.
It is also interesting reading this thread because I have never had a car before where people were so upset over whether they would get upgraded to the latest hardware. I understand people paid for it. I did too. I'm just saying I never had this as an "issue" with any other car.
Yah, there may be regulatory issues with building new cars with recycled hardware. Further, it then opens the possibility that Tesla would need to upgrade the unit in the future, a bigger cost than installing HW3 to begin with.
The Nvidia PX can be used for training, so one option is to add it to their compute cluster (unsure if $/training unit is favorable).
They would need strict control over the units to ensure the quality of the refurb parts, but that would be possible.
More awesomes, more problems...
...and meaningful advantage updates start no earlier than when the Tesla cash crunch is over?
...and meaningful advantage updates start no earlier than when the Tesla cash crunch is over?
Worst case scenario is that we will get it retrofitted after the features can be enjoyed. So the people who got the cars last, will enjoy it first.
They probably can but do they really needs hundreds of thousands of spares laying around?
No other car has promised FSD functionality. People are interested in getting the upgrade they paid for. Some have been waiting nearly 2.5 years.
They can be, but I am pretty confident that they are almost useless for training _large_ neural networks if you already have an AWS system up and running.
Don't think they would be very valuable to the crypto people either. Maybe some ROS-project people would like them for a few more years.
Is it really hundreds of thousands? Tens of thousands yes but hundreds?
Yes, Tesla overpromises. Then again, I keep enjoying that my car just got yet more features I didn't pay for, I was never promised Dog Mode or Sentry Mode. I don't see anything nefarious. I just see Musk as overly optimistic. He sees 80% of traffic lights being recognized and thinks a little more training and they are there. I wish he would realize it is a lot tougher than that.
Yeah, I'm a glass half full type of guy. I have bigger issues than getting bent out of shape about when I'll get HW3. That's just me.
Yeah, not the most elegant, unless you are running AP2 validation and want parallelism. Like the old days when you could combine multiple 1/4 Meg sticks of RAM.
I'd consider ordering a recycled Tesla NN dev kit for work though. Also good for schools, charitable donation?
The key is the same on all the units. They embedd the public version of it in firmware:
Code:
APE_PROD_PUBKEY_HASH="0x94a07b64d0b74bd644852b482db075fbf3c1aaa917958fcb6aca21df123713ac"
APE_PROD_PUBKEY_PATH="/sys/devices/3820000.efuse/public_key"
APE_PROD_SBK_ENC="0x443d412bed681c34226a2caf35094182"
APE_PROD_SBK_PATH="/sys/devices/3820000.efuse/secure_boot_key"Is there a global flash erase on the hardware?
Edit: is the firmware key (decryption) the same as a bootloader key (write protection)?
ROS project people like myself would love them.
That's why it's really unfortunate that I can't see anyway of Tesla being able to re-sell them to us or to give them to colleges and high schools for those kinds of projects.
Tesla could possibly re-use some for side projects like an autonomous warehouse bot to ferry people around the gigafactory.
Hopefully Tesla will find some use for them so they don't end up in a landfill with 700,000 Atari games
Although the Irony in that would be funny considering the Atari easter eggs.
Maybe Tesla should include the ET game as an Easter egg before HW3 upgrades.
All joking aside I do think Tesla will come up with something that we just haven't thought of.
Similar threads
- Article
