Tesla autopilot HW3

[jimmy_d]

So it’s probably been covered before but can you tell me what the hardware differences are currently? I thought it was only the radar and that it was just a different supplier.... also mention of different steering motors.... I am curious what @jimmy_d thinks will happen with A.P 2.0. Oh I miss @jimmy_d analysis..... I know the software is the main thing to achieve FSD and not the hardware ...but I don’t understand how Elon is saying AP2 would be better for a while.... because with the NN training with new silicon? Someone edumacate me

HW2.0 to HW2.5 sensor changes include a different part number for the radar (from a different vendor) and different part numbers for the cameras (from the same vendor). The exact spec on the radar isn't known but a similar part number from the same vendor and a comparison to the wiring harness changes suggest that the new radar has additional signal processing features and enhanced range. The primary change in the driving cameras (not including the backup) seems to be a change in the filter that added an additional color filter (blue) - probably at the expense of a reduced dynamic range for the clear color channel.

The APE (self driving computer) hardware also changed between 2.0 and 2.5 with the biggest change that we know about being the addition of a second Tegra CPU/GPU/SOC. There's a good chance that the presence of this second Tegra is currently enabling features like dashcam which haven't yet been enabled for 2.0 hardware. I haven't seen any credible evidence that the second Tegra currently makes a difference in the ADAS capabilities although it's not unreasonable to think that it could if Tesla chose to take advantage of it. Analysis of APE code looking for those differences has so far failed to turn up anything substantial between 2.0 and 2.5 ADAS operation.

What does this mean for driving capability and in particular the ability to implement FSD? We don't know. There's a lot of speculation and no shortage of people who are convinced that they know the answer, but I am aware of no publicly available objective analysis that would tell us how these changes affect the performance of an ADAS system. If the radar has enhanced specifications then it stands to reason that the improved performance would be helpful, but how helpful and in what particular situations it would be helpful isn't clear. Pundits are fond of analyzing the radar as if it were a backup for AP2's cameras but we don't actually know if it gets used this way. We know that the driving logic is integrating radar and visual signatures for moving objects which fall into the view of both sensors, but vehicles visible only to cameras still get the same set of attributes (centroid, range, velocity) which radar provides. The radar attributes certainly provide additional information that is likely to be useful but we don't know how much difference that additional information makes to overall performance so we can't know how much impact a change in the quality of the radar data would have on overall performance. It's worth noting that there are ADAS systems which are camera only (e.g. Nissan ProPilot) so there's an existence proof for consumer products which do not rely on radar data.

Similarly with the cameras - the tradeoff between the extra color channel and the dynamic range reduction needs to be empirically evaluated, but I haven't seen any research that addresses this topic and I have yet to see an analysis that was so compelling that it carries weight in the absence of empirical data.

Presumably Tesla made these hardware changes with the intention of making FSD easier to implement, but factors such as vendor relationships, component availability, pricing and so forth also play a role and we don't know any of that stuff. I would be inclined to think that HW2.5 is likely more capable that HW2.0 but that is for now an unsupported opinion.

Public statements by Tesla representatives have consistently claimed that 2.0 and 2.5 will both be FSD capable. This includes statements made as recently as yesterday. To date I haven't seen any evidence that existing 2.0 and 2.5 ADAS capabilities vary in any manner that would be relevant to whether FSD is doable or not on either platform.

So - there's no external data with any weight. Tesla has a statement out there - do you trust that statement? My observation of the debate surrounding this issue is that people with strong opinions vary primarily in terms of whether they are inclined to believe Tesla or not. All the arguments I've seen boil down to rationalizations of the author's reason to believe or disbelieve what Tesla is saying.

 

[verygreen]

Is there a global flash erase on the hardware?

there is but it does not help you - these bits are stored in the CPU itself. cannot be erased once ODM fuse is blown

Edit: is the firmware key (decryption) the same as a bootloader key (write protection)?

different bootloader and firmware keys of course. But it does not help you one bit for obvious reasons.

 

[jimmy_d]

So it’s probably been covered before but can you tell me what the hardware differences are currently? I thought it was only the radar and that it was just a different supplier.... also mention of different steering motors.... I am curious what @jimmy_d thinks will happen with A.P 2.0. Oh I miss @jimmy_d analysis..... I know the software is the main thing to achieve FSD and not the hardware ...but I don’t understand how Elon is saying AP2 would be better for a while.... because with the NN training with new silicon? Someone edumacate me

Why would AP2 HW be better right now than HW3? The NN compute capability of HW3 is expected to be far superior to AP2, but the architecture also changes in ways that are not a superset of what AP2 does. Because of this the NN code has to be reimplemented on the new hardware and they may currently be at a point where it doesn't make sense to create bit-for-bit identical implementations of the NN for both hardware platforms. This could be related to how optimizations differ for the two platforms, it could be related to having different numerical library implementations, or it could just be a matter of quantization variation within the dataflow of the hardware. Another factor on the NN side is that regression testing in their datacenter can be done on Nvidia's enterprise class hardware using the same libraries and frameworks as are run in the vehicle, which allows for precise evaluation of the downstream effects of stuff like network compression which have to be done after training but before deployment. Tesla may not yet have datacenter implementations of their NN chip in which case regression testing would require emulation of the final hardware config, or possibly they would ignore the differences in favor of evaluating at a different layer. The overall result is that there would be minor differences which are currently hard to tune out for HW3 - possibly resulting in slightly lower performance in the field. That would change as they extend their infrastructure and refine the development and deployment frameworks and processes.

Also, HW3 differs in ways that will require changes to software that isn't directly NN related. The CPU architecture has changed slightly and it's possible that things like ram, flash, and IO interface devices have changed resulting in some changes being needed to drivers and libraries. To the extent that those parts haven't been polished yet there could be issues which impact performance ore reliability of ADAS operation.

 

[chillaban]

Even beyond architectural differences, remember MCU2 at launch was super unstable and didn’t connect to WiFi, didn’t respond to 2 finger resets, etc.

New platforms can be rough around the edges and Tesla likes to ship things that I am more used to being about 3-6 months from shipping (aka hardware final or too late to change, software and stability still in progress)

 

[ChrML]

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.

They could ship the detached HW2 / HW2.5 units to some place with cooling, then hook them up with power and network through some rack. They are GPU's and they are already paid for, so 1000s together could be used for training their neural networks.

No idea if this would be economical at all tho versus just renting capacity at some rendering-farm. Like they probably do today when they train/iterate their neural networks.

 

[dgatwood]

Only states FSD computer for HW2.5. Says HW2.0 will get FSD. You must read Elon's words in the most narrow way possible.

So what you're suggesting is that HW2.5 users will get HW3 with real FSD, but HW2.0 users will get hand-me-down HW2.5 with really bad FSD?

Hopefully there is some option that neither you or I are seeing because trashing/recycling 100,000 Nvidia PX2 computers is a terrible waste.

How do you get a number that small? There are what, about 500k Teslas out there? About 300k on AP2 and later, give or take? Assuming full self driving works, eventually every Tesla owner will either buy the feature or sell the car to someone who does. Or are you assuming that 200k of the 300k will get repurposed to upgrade the AP1 users for all the features they were promised?

HW2.0 to HW2.5 sensor changes include a different part number for the radar (from a different vendor) and different part numbers for the cameras (from the same vendor). The exact spec on the radar isn't known but a similar part number from the same vendor and a comparison to the wiring harness changes suggest that the new radar has additional signal processing features and enhanced range. The primary change in the driving cameras (not including the backup) seems to be a change in the filter that added an additional color filter (blue) - probably at the expense of a reduced dynamic range for the clear color channel.

That seems unlikely unless if the underlying sensor also changed. The dynamic range of a camera is defined by the image sensor's full well capacity minus any reduction caused by amplification beyond a certain threshold, capped by the number of bits of output. None of those factors are affected by the color filter in any way. (Pedantically, if blue fully replaced the clear channel, it would reduce low-light sensitivity, which would require more amplification and thus reduce dynamic range in low light, but the dynamic range at base ISO would still be identical.)

Changing the sensor's color filter would, of course, result in reduced resolution in the B/W channel.

The APE (self driving computer) hardware also changed between 2.0 and 2.5 with the biggest change that we know about being the addition of a second Tegra CPU/GPU/SOC. There's a good chance that the presence of this second Tegra is currently enabling features like dashcam which haven't yet been enabled for 2.0 hardware. I haven't seen any credible evidence that the second Tegra currently makes a difference in the ADAS capabilities although it's not unreasonable to think that it could if Tesla chose to take advantage of it. Analysis of APE code looking for those differences has so far failed to turn up anything substantial between 2.0 and 2.5 ADAS operation.

I'm pretty sure the original intent was to have it be a hot-failover (and possibly to allow for silent testing of experimental NN model variants in parallel with the active model, for validation purposes). That's why all the sensors feed both sides, and that's why the sensors in AP2.5 all have a secondary set of wiring harnesses all the way back to the computer, IIRC. That's what could make AP2 to AP3 upgrades slightly problematic, though obviously they could just disable the hot failover, albeit at some cost in terms of safety.

 

[verygreen]

How do you get a number that small?

I think you are mis-attributing that quote as it's not mine?

I'm pretty sure the original intent was to have it be a hot-failover

nope. cannot have that. it has no storage at all, also only half the RAM of the main node. Also it does NOT havea dedicated GPU unlike that main node.

Now on HW3 they learned their lesson and it looks like the -b node is fully featured except I think it still has no storage.

 

[electronblue]

So what you're suggesting is that HW2.5 users will get HW3 with real FSD, but HW2.0 users will get hand-me-down HW2.5 with really bad FSD?

No, what @verygreen is suggesting is that HW2 will only get software updates towards FSD while HW2.5 gets HW3 hardware and software updates towards FSD.

This is one rare thing where I think he is — uncharacteristically — entirely mistaken.

 

[electronblue]

Public statements by Tesla representatives have consistently claimed that 2.0 and 2.5 will both be FSD capable. This includes statements made as recently as yesterday. To date I haven't seen any evidence that existing 2.0 and 2.5 ADAS capabilities vary in any manner that would be relevant to whether FSD is doable or not on either platform.

So - there's no external data with any weight. Tesla has a statement out there - do you trust that statement? My observation of the debate surrounding this issue is that people with strong opinions vary primarily in terms of whether they are inclined to believe Tesla or not. All the arguments I've seen boil down to rationalizations of the author's reason to believe or disbelieve what Tesla is saying.

If so, I would consider myself an exception to that rule.

I wholeheartedly disbelieve what Tesla is saying in general but I also believe AP2 will get the HW3 upgrade (assuming HW2.5 gets it too) because I don’t see a realistic or reasonable rationale towards the opposite — that’s because these things are not a matter of any sort of blind belief in the end but of rational analysis of what is going on.

With rational analysis, there is reason to disbelieve Tesla in general but also rational reason to believe them in particular instances like this one.

Now, whether or not you believe any HW2/2.5 cars will get retrofits or how long that will take, that is a much harder question to answer.

 

[mongo]

No, what @verygreen is suggesting is that HW2 will only get software updates towards FSD while HW2.5 gets HW3 hardware and software updates towards FSD.

This is one rare thing where I think he is — uncharacteristically — entirely mistaken.

@verygreen is pessimistic on Elon Tweets, which I can sympathize with. However:

(This does leave sensor upgrades as an unknown)

 

[boonedocks]

@verygreen is pessimistic on Elon Tweets, which I can sympathize with. However:
View attachment 392148
(This does leave sensor upgrades as an unknown)

Could be that HW2.0 will get more than just a computer....Elon really does craft words in a purposefully confusing obtuse way. I think he believes he is being super clear but ends up being vaguely muddy. He really needs a wordsmith to help with his social media communications.

His “Anyone who bought FSD will get it” was the most clear he has been in a while about it.

As a February 2017 HW2.0 pre-paid FSD at delivery owner I am more comfortable with the whole situation than I have ever been. (That and some very encouraging emails from engineering LOL). My only concern is that my car at 68k mikes will have to have mini overhaul by the time it is retrofitted. :/

 

[fasteddie7]

There's us nerds that dive into the code, and there are regular people who are hesitant to upgrade their phones software because it "changes things" and they don't want to relearn. I think a lot of us are looking at this like everyone is like us. When an average person goes in and is told they are getting an computer upgrade or can get one for several thousand dollars they will decline. Tesla may be counting on those owners who are still looking at Tesla like a traditional automaker and those feature upgrades come when they buy new cars. As for the rest of us, it's going to be like AP1 where the mention of it slowly fades away in favor of some new shiny thing dangled in front of us to make us forget because we can't resist new tech, likely a refresh on interior and exterior for me. I'd imagine those that jumped on back in 2016 when the "Tesla Network" would drop you off and make money for you picking up passengers while you were at work on the button you clicked to purchase on the website, and the 3 month maybe, 6 months definitely talk that faded away might feel burnt and shy away from purchase, I for one will wait for the next one to have the abilities I want before I purchase and get excited about new features that are thought up on the fly like sentry mode and dog mode like I get excited about promised features now.

 

[electronblue]

Could be that HW2.0 will get more than just a computer....

I have never seen any reason to believe that it would. I don’t get where this pipedream comes from.

It will be a computer swap as Tesla has said.

I could see them swapping some fuses or somesuch small supporting techncial bits if the retrofit absolutely needs it, but I don’t see them changing sensors.

 

[mongo]

I have never seen any reason to believe that it would. I don’t get where this pipedream comes from.

It will be a computer swap as Tesla has said.

I could see them swapping some fuses or somesuch small supporting techncial bits if the retrofit absolutely needs it, but I don’t see them changing sensors.

If you squint while reading the tweet. All FSD get computers, hw2.5 only needs computer. That leaves the possibility that hw2.0 needs more than computer. The cameras and radar could be replaced (even if it requires adapters). However, it may not be needed.

 

[BigD0g]

Could be that HW2.0 will get more than just a computer....

I see very little to no chance of this happening, they'll give us the AP3 computer to avoid the lawsuits "someday", but they are going to give us new wiring harnesses and cameras and radars, to much work and they never agreed to do anything like that. All they've ever really said was the CPU upgrade is needed. That being said, they did say "all hardware needed for FSD" was already on the car, so as soon as a sensor on > 2.5 becomes "needed" for FSD, well then the lawyers will probably get involved...

 

[MasterchiefTS]

I want to see an demo of Hardware 3.

 

[mongo]

I see very little to no chance of this happening, they'll give us the AP3 computer to avoid the lawsuits "someday", but they are going to give us new wiring harnesses and cameras and radars, to much work and they never agreed to do anything like that. All they've ever really said was the CPU upgrade is needed. That being said, they did say "all hardware needed for FSD" was already on the car, so as soon as a sensor on > 2.5 becomes "needed" for FSD, well then the lawyers will probably get involved...

It could be a chunk of work, but if sales are up, HW3 is built into new cars, and FSD development is stable, then they will have crazy amounts of cash coming in from new purchases of AP/ FSD that would more than cover upgrade costs.

 

[diplomat33]

I wonder if maybe one of the benefits of AP3 will be a better display. Right now, on AP2/2.5, we sometimes get the "dancing cars" on the display. I wonder if that issue will be solved with AP3. Also, since AP3 can process data faster and better, I also wonder if AP3 will be able to give us more info as well, where the display actually will show us intersections,traffic lights, stop signs, and cross traffic etc...

 

[Bet TSLA]

I wonder if maybe one of the benefits of AP3 will be a better display. Right now, on AP2/2.5, we sometimes get the "dancing cars" on the display. I wonder if that issue will be solved with AP3. Also, since AP3 can process data faster and better, I also wonder if AP3 will be able to give us more info as well, where the display actually will show us intersections,traffic lights, stop signs, and cross traffic etc...

It could, but why? As humans become less responsible for driving, there is less reason to provide them with any information at all.

 

[diplomat33]

It could, but why? As humans become less responsible for driving, there is less reason to provide them with any information at all.

By that logic, why even bother showing us the blue lane lines or cars around us? You are right that a FSD car would not need to give us that info but right now, since our cars are not FSD yet, I think showing us the information is useful to give us more confidence in AP.