This is expected if they are separate nodes, though, that communicate by gige (as on the diagram), it was already discussed.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
AP2.0 Cameras: Capabilities and Limitations?
- Thread starter lunitiks
- Start date
How does discussing mobileye which is a big part of tesla history equates to shilling and no they have a L2 system in production.
Tesla and Elon has made outlandish claims like these every time and have failed to deliver. The only thing in customers hands is a L2 system. Infact it was only like last month that partial payment from EAP were recognized. FSD payments haven't even been recognized yet.
Others have, Volvo for example went from pilot Assist 1 to 2 with a software update.
Unlikely because tesla actually said: "That said, Enhanced Autopilot should still be considered a driver's assistance feature with the driver responsible for remaining in control of the car at all times."
I think what alot of people are confused about is what EAP actually is. Alot will be surprised that Its not based on FSD code development, for lane changing on highway for example, its simply replacing the usage of ultrasonics with the two rear cameras. EAP is only meant to accomplish the promises of AP1 that never happened. Smart Summon will also do what summon was promised to do and also on-ramp/off-ramp, etc. EAP is literally a glorified AP1.
stopcrazypp
Well-Known Member
Beat me to it. Yes, it's expected if they are separate nodes. Another thing I should add is that if "lb" really stands for "load balancer", the two nodes may even be practically identical.
no, I am practically convinced lb is some sort of a dumb gps thingie.
I plan to research it in more detail real soon now
stopcrazypp
Well-Known Member
Yes, figuring what lb is would really help. The only thing I'm worried about in terms of going the software route to figure the configuration out is if Tesla left the other SoC unused for EAP, then that may not show up in the software.
But looking at the physical board doesn't have the same problem.
Dang this thread just exploded while I was sleeping 
According to the Parts Manual, they did update the antenna, though. I suspect it is a Hirschmann Car Communications mouse of some sort.
I totally agree with @bjornb here.
First off, one needs to understand the difference between the "Autocruise" and "Autochauffeur" boards:
"Autocruise" only consists of a "Parker" chip, which has 6 CPUs (2 x Denver + 4 x Coretex A57) and an integrated GPU (iGPU). This chip is also known as Tegra X2 or T186.
"Autochauffeur" has that times two (i.e. two "Parker" chips, totalling 12 CPU cores and 2 iGPUs). ADDITIONALLY it has two discrete GPUs (dGPUs) -- probably MXM 3.1s -- mounted on the flip side of the board. Each of these two dGPUs is accompanied by dedicated video memory @4GB (Elpida W4032BABG-60-F). Check this out: Need for Speed. In conclusion: "Autochauffeur" (the so-called full board) is not merely a double "Auocruise" board - it's far more than that because of the dGPUs.
The enclosure pictures taken by me, @bjornb and others show that the board is mounted almost flush with the bottom of the steel/aluminium enclosure. Not much room for dGPUs on the bottom!
Certainly Tesla has a very modified version of the PX2, if it really can be called PX2 at all. Elon certainly didn't mention PX2 in his press conference speech. Yet Nvidia has boasted that it is indeed a PX2 in there.
In AP2-cars, the GPS-signal from the antenna in the trunk lid terminates directly in the AP2ECU ("PX2") on a dedicated connector. How the "PX2" processes and sends this to the MCU is a big mystery thus far. All I know is that the Ublox device in the MCU is a LEA-6R-0-001 in AP1-cars. (BTW in AP1-cars, the antenna cable terminates directly in the MCU.) That said, the MCU was not updated with AP2, so it should be the same GPS-module. The latest code dump from @verygreen mentions Ublox:
According to the Parts Manual, they did update the antenna, though. I suspect it is a Hirschmann Car Communications mouse of some sort.
I totally agree with @bjornb here.
First off, one needs to understand the difference between the "Autocruise" and "Autochauffeur" boards:
"Autocruise" only consists of a "Parker" chip, which has 6 CPUs (2 x Denver + 4 x Coretex A57) and an integrated GPU (iGPU). This chip is also known as Tegra X2 or T186.
"Autochauffeur" has that times two (i.e. two "Parker" chips, totalling 12 CPU cores and 2 iGPUs). ADDITIONALLY it has two discrete GPUs (dGPUs) -- probably MXM 3.1s -- mounted on the flip side of the board. Each of these two dGPUs is accompanied by dedicated video memory @4GB (Elpida W4032BABG-60-F). Check this out: Need for Speed. In conclusion: "Autochauffeur" (the so-called full board) is not merely a double "Auocruise" board - it's far more than that because of the dGPUs.
The enclosure pictures taken by me, @bjornb and others show that the board is mounted almost flush with the bottom of the steel/aluminium enclosure. Not much room for dGPUs on the bottom!
Certainly Tesla has a very modified version of the PX2, if it really can be called PX2 at all. Elon certainly didn't mention PX2 in his press conference speech. Yet Nvidia has boasted that it is indeed a PX2 in there.
Neither can I, that's why I started this thread allready back in DecemberI can't wait for the first person to post photos of the board.
Last edited:
stopcrazypp
Well-Known Member
Dang this thread just exploded while I was sleeping
In AP2-cars, the GPS-signal from the antenna in the trunk lid terminates directly in the AP2ECU ("PX2") on a dedicated connector. How the "PX2" processes and sends this to the MCU is a big mystery thus far. All I know is that the Ublox device in the MCU is a LEA-6R-0-001 in AP1-cars. (BTW in AP1-cars, the antenna cable terminates directly in the MCU.) That said, the MCU was not updated with AP2, so it should be the same GPS-module. They did update the antenna, though.
I totally agree with @bjornb here.
First off, one needs to understand the difference between the "Autochauffeur" and "Autocruise" boards:
"Autocruise" only consists of a "Parker" chip, which has 6 CPUs (2 x Denver + 4 x Coretex A57) and an integrated GPU (iGPU). This chip is also known as Tegra X2 or T186.
"Autochauffeur" has that times two (i.e. two "Parker" chips, totalling 12 CPU cores and 2 iGPUs). ADDITIONALLY it has two discrete GPUs (dGPUs) -- probably MXM 3.1s -- mounted on the flip side of the board. Each of these two dGPUs is accompanied by dedicated video memory @4GB (Elpida W4032BABG-60-F). Check this out: Need for Speed. In conclusion: "Autochauffeur" (the so-called full board) is not merely a double "Auocruise" board - it's far more than that.
The enclosure pictures taken by me, @bjornb and others show that the board is mounted almost flush with the bottom of the steel/aluminium enclosure. Not much room for dGPUs on the bottom! Also, where's the cooling for these?
Certainly Tesla has a very modified version of the PX2, if it really can be called PX2 at all. Elon certainly didn't mention PX2 in his press conference speech. Yet Nvidia has boasted that it is indeed a PX2 in there.
Neither can I, that's why I started this thread allready back in December
This is really hard to visualize without pictures. I decided to google the actual physical pictures of the boards.
This is the PX2 Autocruise (single Parker; unfortunately I couldn't find a picture of bottom side):
This is the "Autochauffeur" (2 Parker, 2 discrete Pascal GPUS):
The bottom picture I haven't seen previously, but it shows the two discrete Pascal GPU boards connected with MXM connectors to the bottom.
This description would match neither:
Rather that suggests a configuration with 1 Parker and 1 discrete GPU, which Nvidia had never shown. However, this would match up with Elon saying 12 Trillion operations per second (vs 24 DL TOPs quoted by Nvidia for the PX2 Autochauffeur).
Last edited:
You're certainly right about that @stopcrazypp. I think people are up for a surprise when that teardown happens.
Here's some more pics of the dGPUs on the dev board ("bottom" side):
Here's some more pics of the dGPUs on the dev board ("bottom" side):
Last edited:
Not quite
Teslas I/Os are totally different from Nvidias dev board. Tesla uses a combination of Rosenberger 59Z115 (double connectors), Rosenberger 59Z113 and D4Z001 (single connectors), plus a couple of Japanese(?) connectors like Sumitomo 6098-5611 and 6098-4008. Furthermore, the Tesla unit does not have a OBDII-connector, SD-slot or the "double decker" USB-port."Counting I/Os" actually gets you nowhere. According to this source (see p. 6), the Autocruise variant is able to "connect & fuse data from up to 8 cameras", plus radar and lidar. I've seen pictures of the Autocruise with two and three camera connectors.
Teslas AP2ECU is custom indeed
Not quite
"Counting I/Os" actually gets you nowhere. According to this source (see p. 6), the Autocruise variant is able to "connect & fuse data from up to 8 cameras", plus radar and lidar. I've seen pictures of the Autocruise with two and three camera connectors.
Teslas AP2ECU is custom indeed
Did any of your pics have Nvidia's SKU on them? Can see a sticker on one but cant read it.
On your pictures there's an odd space next to usb-c that looks like sdcard slot that was just not put in placeNot quite
The sticker just tells you Teslas part no (1078321-XX-X).
Speaking of which: The AP2ECU has been revised a couple of times since october '16. Current version is 1078321-70-G ("can replace all old versions, 1078321-70-C and 1078321-70-D").
Yes, I agree that customers have a L2 system now and EAP is also L2, but what is the potential of the system in your opinion?
L2? L3? L4? L5?
Lets assume for now that the processing power of the autopilot ECU is not a limitation.
Elon Musk has said 'all cars exiting the factory have hardware necessary for Level 5 Autonomy' and 'with Level 5 literally meaning hardware capable of full self-driving for driver-less capability' (but I have a feeling you have a meaning about Elons credibility.. ;-) ).
On their own Autopilot page they actually do not go that far:
" the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver".
"enabling full self-driving in almost all circumstances, at what we believe will be a probability of safety at least twice as good as the average human driver".
To me it seems like they say the potential is L4 on the web page(?)
For reference, posted by @Bladerskb earlier:
About processing power:
The Tesla web page says "Processing Power Increased 40x".
On the AP2 presentation Elon said "The compute power increases by a factor of 40 (40x increase in compute power), it’s such a gigantic increase in computing power, infact the computer is capable of 12 trillion operations per second".
Do we know anything more specific about this?
What kind of 'operations per second' is he talking about? TFLOPS? DLTOPS?
I have seen EyeQ3 described with a performance of 0,2 and 0,256 TFLOPS, but I dont know whether this is a comparable value to what is presented about the Nvidia PX2 with 8 TFLOPS/24 DLTOPs (I do sometimes follow and read news regarding new CPUs/GPUs on Anandtech etc, and I know that the values differ whether it is FP32, FP16 TFLOPS etc..,).
The Lea 6R makes sense as it is dead reckoning for tunnels etc., but not raw data (RTK capable) enabled by default. plus only single constellation (GPS) There is a custom firmware hack around to get raw data ( pseudorange, doppler and phase) out on most 6 series Ublox receivers. However the 6 series ublox is now quite old technology, now superceded by the 8 series giving it Glonass, Galileo plus Beidou capability in addition to GPS. this doubles or triples the number of available satelites hence improves urban canyon performance.
From the Ublox site "u-blox 6 modules are previous generation devices. EVA-M8E, NEO-M8L or NEO-M8U are recommended for new dead reckoning designs."
So it would be dissapointing if they are still fitting the LEA-6R.
@arwooldridge Actually the description of the new (AP2) GPS-antenna in the parts manual states "GNSS", as opposed to the old antenna which states "GPS". So I guess it would make sense that they updated the ublox module at some point
Reciprocity
Active Member
Awesome info on this thread but I'm gonna toss a wrench in like I am known to do. Even I have made the mistake of thinking Tesla could use Xavier in the future, but then I remembered they have a deal with Samsung to make a chip. At first i assumed this was for the mpu? If that's what it's called. But then I thought about how Apple made its own are processors for the iPhone and iPad.
They could be licensing something from Nvidia but I think the goal would be to design a chip that only has what Tesla needs and not as full featured, AKA expensive, as it relates to the px2/Xavier. No lidar support for example.
The goal would be two fold, one processor for everything that is built specifically for Tesla and can be iterated over time to support upgrades of existing cars and future cars. Price and power consumption minimized. Support for augmented reality HUD? Tesla won't want several disparate systems and cpus that drive up complexity and cost.
Now what to do with 100,000 drive px2 boards? Can they rack then up and build a machine learning super computer in a data center?
Sorry to derail this thread again, but I think it's the right audience, minus one guy who won't be named.
They could be licensing something from Nvidia but I think the goal would be to design a chip that only has what Tesla needs and not as full featured, AKA expensive, as it relates to the px2/Xavier. No lidar support for example.
The goal would be two fold, one processor for everything that is built specifically for Tesla and can be iterated over time to support upgrades of existing cars and future cars. Price and power consumption minimized. Support for augmented reality HUD? Tesla won't want several disparate systems and cpus that drive up complexity and cost.
Now what to do with 100,000 drive px2 boards? Can they rack then up and build a machine learning super computer in a data center?
Sorry to derail this thread again, but I think it's the right audience, minus one guy who won't be named.
Similar threads
- Article
