Neural Networks

[chillaban]

I don't think they will update all the cars before sending out HW3 NN. They will start sending HW3 NN to atleast some % of cars as the work to port it finishes. Then, they will upgrade all the cars over 2 or 3 quarters.

If they do this right, by the time NOA on city roads is ready for fleetwide deployment - sometime next year, all the upgrades would be over.

I honestly think they built a small number of HW3 cars just to seed the fleet and flesh out the hardware. It’s likely not truly ready yet.

Most silicon goes through 2+ revisions about 6-8mo apart from each other before they’re ready for customers. Wouldn’t be surprised of HW3 will need to be retrofitted too when HW3 Rev B comes out

 

[EVNow]

I honestly think they built a small number of HW3 cars just to seed the fleet and flesh out the hardware. It’s likely not truly ready yet.

Most silicon goes through 2+ revisions about 6-8mo apart from each other before they’re ready for customers. Wouldn’t be surprised of HW3 will need to be retrofitted too when HW3 Rev B comes out

We don't know how many cars they built with Rev 1 of HW3 - and may be now are on Rev 2/3. Its very unlikely they decided to mass produce HW3 without testing thoroughly. Their chip designer doesn't look like an amateur

All cars being produced now (from a couple of months, 4/12 according to EM) have HW3.

 

[Eno Deb]

All cars being produced now (from a couple of months, 4/12 according to EM) have HW3.

There is some evidence that this is not actually the case:

HW2.5 even after April 12th build date

 

[EVNow]

There is some evidence that this is not actually the case:

HW2.5 even after April 12th build date

There may be some with HW 2.5 (or the API is returning wrong version) - but that is not the same as what the poster I replied to claimed. They thought only a small # of cars have HW3 and rest have HW2.5. In the last 2 months they have probably built 60k cars now with HW3.

ps : Well, after reading the thread more, apparently quite a few think they have HW2.5. Will continue to monitor.

 

[electronblue]

There is some evidence that this is not actually the case:

HW2.5 even after April 12th build date

There may be some with HW 2.5 (or the API is returning wrong version) - but that is not the same as what the poster I replied to claimed. They thought only a small # of cars have HW3 and rest have HW2.5. In the last 2 months they have probably built 60k cars now with HW3.

ps : Well, after reading the thread more, apparently quite a few think they have HW2.5. Will continue to monitor.

Well to be fair, they said all cars build ”today” April 22nd would have HW3. Maybe they switched back on the 23rd.

This is Musk's tweet from April 22:
Tesla on Twitter
All Tesla cars being built today have the hardware necessary for full self-driving:
8 vision cameras, 12 ultrasonic sensors, radar, and this custom-designed beast of a Full Self-Driving Computer

I think it speaks tons to the lack of perceived transparency and ethics at Tesla that we are no longer even surprised by the possibility that they could make a silent switch though.

Could this be related to the China tariffs, weren’t they related to the FSD Computer?

 

[AlanSubie4Life]

Most silicon goes through 2+ revisions about 6-8mo apart from each other before they’re ready for customers. Wouldn’t be surprised of HW3 will need to be retrofitted too when HW3 Rev B comes out

Another rev is always possible, but they are already on B0 silicon (see 1:13:00 or so on Autonomy Day presentation for development timeline); they taped out alpha silicon August 2017.

Of course, they are also working on the next HW version already: “doing all the obvious things”, whatever that means, but that is a different hardware. If they can’t fulfill promises with HW3, sure, that would need to be retrofitted.

 

[BioSehnsucht]

China tariffs won't be reason for 2.5 vs 3. They all come from the same place.

There is likely only one revision of HW3 in the wild right now. Yes, they would have had several chip and possibly board level revisions before putting it in cars, but they would do that testing before hand. They're not going to be doing multiple revisions and fixes with production vehicles because that would mean massive labor costs to get them all on the final "good" version.

As for HW4 vs HW3, expect it to be focused on cost reduction (via smaller process node) and power reduction (both process node and engineering efficiency improvements to reduce power usage). It will likely get a performance bump just because it's there to be had, but I would be surprised if it had a large performance change of HW3. There may also be updates to other aspects of the design (i.e., the ARM and GPU cores they use may get newer, more efficient versions in HW4) that aren't part of the NN design itself. Similarly, there may be board level improvements for HW4 to reduce costs/power, but don't expect a bunch of HW3 iterations chasing these things unless there's some low hanging fruit (and then it would still be functionally 100% compatible, just different PCB layout and components for reduced costs - don't expect NN chip revisions before HW4 NN chip because those are expensive)

 

[electronblue]

@BioSehnsucht Plausible. One more idea though: Tariffs could still play a role if there is a price difference in the end due to either lower overall AP 2.5 price or existing stock stateside. After all, this is the quarter where Tesla saves with toilet paper.

By the way do I recall wrong or was there talk or even pics of two HW3 revisions already out there, one with the debugging display port (like some older B0s) and one later one with it removed (Bx something was it?)?

 

[ChrML]

I honestly think they built a small number of HW3 cars just to seed the fleet and flesh out the hardware. It’s likely not truly ready yet.

Most silicon goes through 2+ revisions about 6-8mo apart from each other before they’re ready for customers. Wouldn’t be surprised of HW3 will need to be retrofitted too when HW3 Rev B comes out

They confirmed at the autonomous conference that HW3 did have atleast one iteration before it went out. Current mass-produced HW3 is on revision 2 or 3.

 

[Matias]

If they start deploying additional functionality before all the people who have paid for FSD and have been promised a hardware upgrade actually get their upgrade, there will probably be a huge sh!tstorm. My guess is they'll wait until all or most of the upgrades have been done, which will probably not be this year.

What? They are not doing any hardware upgrades at the moment.

 

[Saghost]

What? They are not doing any hardware upgrades at the moment.

AFAIK that's correct, and it certainly would be easier to start by rolling it out to the cars with the latest production hardware and make sure everything works as expected before spending a bunch of money refitting cars, but I can certainly see folks who paid full price in 2016 being really annoyed that people who bought the car three or four years later got the features before they did.

Not sure how Tesla is going to play this.

 

[ChrML]

but I can certainly see folks who paid full price in 2016 being really annoyed that people who bought the car three or four year

Well, the 2016 buyers atleast had the option to buy it back then. Other car manufacturers it would not be possible to buy at all before it's ready.

 

[el_pinguino]

Does anybody know if Tesla's NN model the same globally (true general purpose solution)? If so, are there plans for country-specific localization? Considering the differing traffic laws, road signs, road markings, driver etiquette, etc. etc, I would think country-specific NN models is the way to go to accelerate progress.

 

[mblakele]

Does anybody know if Tesla's NN model the same globally (true general purpose solution)? If so, are there plans for country-specific localization? Considering the differing traffic laws, road signs, road markings, driver etiquette, etc. etc, I would think country-specific NN models is the way to go to accelerate progress.

That would make sense. As I understand it, commercial NN-based systems like Siri, Alexa, etc use a different net for each region they support (American English, British English, Mexican Spanish, European Spanish, etc). A vehicle driving from one country into another could request new blob(s) as soon as navigation shows that it'll cross the border, or when GPS shows that it's likely to cross.

 

[EVNow]

That would make sense. As I understand it, commercial NN-based systems like Siri, Alexa, etc use a different net for each region they support (American English, British English, Mexican Spanish, European Spanish, etc). A vehicle driving from one country into another could request new blob(s) as soon as navigation shows that it'll cross the border, or when GPS shows that it's likely to cross.

Its possible they will do it by regions, rather than countries (esp. for EU).

 

[CarlK]

That would make sense. As I understand it, commercial NN-based systems like Siri, Alexa, etc use a different net for each region they support (American English, British English, Mexican Spanish, European Spanish, etc). A vehicle driving from one country into another could request new blob(s) as soon as navigation shows that it'll cross the border, or when GPS shows that it's likely to cross.

Elon when he was talking about Tesla no longer needs to rely on precision GPS and HD map also said GPS will mostly be used to manage NN subsystems. It could be country to country, region to region and I see no reason why it could not use different NN for different states/cities where there could be vastly different driving habits of drivers there. You shouldn't expect driving in New York is the same as driving in Kentucky or vise versa. NN local training should take care of this issue easily, which incidentally is not that easy take advantage of by the Lidar/HD map approach everyone else is using.

 

[el_pinguino]

NN local training should take care of this issue easily, which incidentally is not that easy take advantage of by the Lidar/HD map approach everyone else is using.

Why not? I'm not aware of any serious autonomous vehicle program that doesn't also employ machine learning.

 

[el_pinguino]

Does anybody know if the Tesla hardware 3 computer has audio input? If not to enhance capability, It may be an eventual regulatory requirement for AVs to, at least, be able to detect and respond to emergency vehicle sirens.

 

[Matias]

Well, the 2016 buyers atleast had the option to buy it back then. Other car manufacturers it would not be possible to buy at all before it's ready.

there’s a legal term for selling something that doesn’t exist.

 

[CarlK]

Why not? I'm not aware of any serious autonomous vehicle program that doesn't also employ machine learning.

Tesla is the only major players, or among any players, that uses deep learning neural net. It requires a huge amount of data to produce enough accuracy. No one else has large fleet like Tesla to implement that. Tesla's NN trained in different regions will fit driving habit of that region the best. Others rely on Lidar, high precision GPS and HD map and is very hard for them to program in the difference between drivers in California and Oregon.