Elon: "Feature complete for full self driving this year"

[S4WRXTTCS]

Where did you read of that sequence? The blog post says:

A reddit post of someone who is supposedly in the early access program.
Introducing a More Seamless Navigate on Autopilot : teslamotors

It's kinda one of those "that's not really what I thought it would be, but okay". moments.

 

[dgatwood]

Not only that, 2000x from 200x frames of AP2 is 10x increase in power. Keep in mind that AP2 consisted of around ~8-10 TOPS. Putting AP3 at 80-100 TOPS. Which is further confirmed through analysis by green.

Not necessarily. They're massively downsampling every frame on ingest just to get 200 fps right now. I doubt they are downsampling with the new hardware. So it is probably quite a bit more than an order of magnitude difference. The only thing we can be reasonably certain of is that it is probably at least 10x and less than... say 50x.

Uhm wow are you serious? There really need to be a qualification to post in this forum. Apparently you and ArkInvest have no clue that a Board contains multiple chips, that there is no one chip that has 120 TOPS. So ArkInvest and tasha ignorantly believes that Tesla's chip is 120TOPS and that Xavier is 30TOPS therefore it will take Nvidia aleast three years to catch Tesla. Thus going around repeating that Tesla a 3+ year lead in hardware.

Actually, if the estimates here are correct, then it might provide O(250) TOPS per chip (50 TOPS at 400 MHz, with a maximum clock speed of 2 GHz), in which case this is seriously easing Xavier's lunch... by a factor of ten per chip, times potentially four chips per board.

That said, at this point, the data we have is so limited that we might as well just be throwing darts at a wall somewhere. I hope we'll get more details in a couple of weeks.

 

[electronblue]

Man you're on the ball as that's the first I heard of that blog post.

I didn't even realize my car had a steering wheel capable of vibrating. That's good news for us, and sadly more bad news for AP2 people.

Are we talking of ”AP2.1 really” that is nearly the same as AP2 according to Tesla back in the summer of 2017... nothing to see here...

Given the steering wheel question though that is an interesting one as I believe all Autopilot steering wheels are capable of vibrating. I wonder what caused this one?

 

[S4WRXTTCS]

Are we talking of ”AP2.1 really” that is nearly the same as AP2 according to Tesla back in the summer of 2017... nothing to see here...

Given the steering wheel question though that is an interesting one as I believe all Autopilot steering wheels are capable of vibrating. I wonder what caused this one?

After I wrote that it dawned on me that yes in fact I had a steering wheel vibration on my AP1 car. I did for like a day or two before I realized what was causing that god awful noise, and weird vibration in the steering wheel. It was the LDW system.

I completely forgot about it.

I've never had it on in the Model 3 I have now so I've never experienced the steering wheel vibration.

I have no explanation as to why HW2 won't have the steering wheel vibration option that HW2.5 gets. I'm not aware of any differences as it pertains to it.

 

[electronblue]

@S4WRXTTCS A demand lever perhaps.

 

[Bladerskb]

Not necessarily. They're massively downsampling every frame on ingest just to get 200 fps right now. I doubt they are downsampling with the new hardware. So it is probably quite a bit more than an order of magnitude difference. The only thing we can be reasonably certain of is that it is probably at least 10x and less than... say 50x.




Actually, if the estimates here are correct, then it might provide O(250) TOPS per chip (50 TOPS at 400 MHz, with a maximum clock speed of 2 GHz), in which case this is seriously easing Xavier's lunch... by a factor of ten per chip, times potentially four chips per board.

That said, at this point, the data we have is so limited that we might as well just be throwing darts at a wall somewhere. I hope we'll get more details in a couple of weeks.

No we are not throwing darts at the wall. Elon said it can run AP2 NN at 2000 frames versus AP2 running it at 200 frames. This isnt full resolution as they are running the same model with the exact same input to come up with this number.

It's two chips clocked at 400 MHZ, there is no max clock rate of 2 ghz. Just running your chip 100 more mhz than what its thermal load and chip is designed for will cause it to be unstable, overheat, shutdown and damaged, let alone something stupid as 2GHZ. Stop making stuff up and stop posting BS.

 

[mongo]

It's two chips clocked at 400 MHZ, there is no max clock rate of 2 ghz. Just running your chip 100 more mhz than what its thermal load and chip is designed for will cause it to be unstable, overheat, shutdown and damaged, let alone something stupid as 2GHZ. Stop making stuff up and stop posting BS.

And you somehow know what the design speed is? From the link

Super late edit - I looked into the DTB for the device (something I should have done from the start) and the CPU cores could go up to 2.4GHz, the TRIP devices up to 2GHz it looks like? (the speeds quoted initially are from bootloader).

Or are you making stuff up?

 

[Bladerskb]

And you somehow know what the design speed is? From the link

Or are you making stuff up?

The clock speed during operation wont fluctuate. It's set. Period. You can try to set any chip to run at whatever clock speed. That doesnt mean it wont shutdown immediately.

The clock speed outlined matches with the thermal load you would typically want in a chip like that. And it also matches with Elon comparison performance statement with ap2. Anything else is making stuff up

 

[WillK]

Man you're on the ball as that's the first I heard of that blog post.

I didn't even realize my car had a steering wheel capable of vibrating. That's good news for us, and sadly more bad news for AP2 people.

I have AP2 and the steering wheel does vibrate for lane Departure Warning. But there does seem to be a difference in this press release.

 

[mongo]

The clock speed during operation wont fluctuate. It's set. Period. You can try to set any chip to run at whatever clock speed. That doesnt mean it wont shutdown immediately.

The clock speed outlined matches with the thermal load you would typically want in a chip like that. And it also matches with Elon comparison performance statement with ap2. Anything else is making stuff up

It is set by the DTS/DTB file which is what was referenced. There is also no reason they could not have a sentry mode speed and an FSD speed for power savings. Further, you do not know the level of cooling the chip has, nor its actually power dissipation.

On top of that, input image size and frame rate are not the main determiners of processing power needed (though they do scale), NN size is. If I give my NN twice as many layers (doubling coefficients), I need twice the processing power for the same image rate.

So unless you start providing some real data to support your position, I fear it is you who are making things up...

 

[Bladerskb]

It is set by the DTS/DTB file which is what was referenced. There is also no reason they could not have a sentry mode speed and an FSD speed for power savings. Further, you do not know the level of cooling the chip has, nor its actually power dissipation.

On top of that, input image size and frame rate are not the main determiners of processing power needed (though they do scale), NN size is. If I give my NN twice as many layers (doubling coefficients), I need twice the processing power for the same image rate.

So unless you start providing some real data to support your position, I fear it is you who are making things up...

You Tesla fans always coming up with ridiculous theories. 3 years ago it was there is no need for a chip upgrade. Tesla and Elon says this is enough. Elons team is so good they can create efficient software which nvidia cant, meaning they wont need any additional compute. Plus nvidia main goal is to sell more compute so they will exaggerate what's needed.

Fast forward 3 years, it's the same BS but different spin. Now it's we have the best chip but we dont have all the info so it might be 1000tops. Who knows?


You Tesla fans never cease to amuse me.

On top of that, input image size and frame rate are not the main determiners of processing power needed (though they do scale), NN size is. If I give my NN twice as many layers (doubling coefficients), I need twice the processing power for the same image rate.

So unless you start providing some real data to support your position, I fear it is you who are making things up

If your layers change then you have a completely different network. Which you wont use to compare performance of chips. You wouldn't run one chip with VGG16 and another with GoogleNET and then compare their performance. Or a FCN vs one without a FC Layer.

Elon already told us how fast the hardware is after their internal comparison, you dont show how fast your chip is compared to its predecessor by running far more intensive network.

Typical Tesla fans resorting to making illogical statements to console themselves that their new hardware is inferior to basically every NN processor made today.

Almost 3 years ago I had ppl debating with me how the ap2 hardware was more than enough, maybe even too much power, now where are they?

 

[Randy7fx]

@Bladerskb - thought you said you were taking a break until April 19th? Perhaps this is the first time in the history of TMC that you were wrong??

 

[OPRCE]

On Apr.19th I want an “investor” to ask Karpathy if the HW3 with FSD will safely and reliably stop from 90mph for stationary vehicles on the highway and when we can expect that HW2.5 vehicles with AP will do the same?

This is the top feature I want to see in the Tesla software, namely a toggle to switch Firetruck Super-Destruction mode *off*.

 

[banned-66611]

Elon said it can run AP2 NN at 2000 frames versus AP2 running it at 200 frames.

Elon Elon Elon. When are you going to learn that he says any old *sugar* and most of it isn't true?

Now he's down to gaslighting you with fake FSD. Come on buddy, if he was your husband it would be an abusive relationship.

 

[mongo]

You Tesla fans always coming up with ridiculous theories. 3 years ago it was there is no need for a chip upgrade. Tesla and Elon says this is enough. Elons team is so good they can create efficient software which nvidia cant, meaning they wont need any additional compute. Plus nvidia main goal is to sell more compute so they will exaggerate what's needed.

Fast forward 3 years, it's the same BS but different spin. Now it's we have the best chip but we dont have all the info so it might be 1000tops. Who knows?


You Tesla fans never cease to amuse me.

Do you or do you not have hard data that contrasts to verygreen's reading of the DTS file (you know, the one that sets the hardware operating parameters?)

If your layers change then you have a completely different network. Which you wont use to compare performance of chips. You wouldn't run one chip with VGG16 and another with GoogleNET and then compare their performance. Or a FCN vs one without a FC Layer.

Elon already told us how fast the hardware is after their internal comparison, you dont show how fast your chip is compared to its predecessor by running far more intensive network.

Typical Tesla fans resorting to making illogical statements to console themselves that their new hardware is inferior to basically every NN processor made today.

Almost 3 years ago I had ppl debating with me how the ap2 hardware was more than enough, maybe even too much power, now where are they?

https://www.fool.com/earnings/call-transcripts/2018/10/25/tesla-motors-inc-tsla-q3-2018-earnings-conference.aspx

Elon R. Musk -- Co-Founder, Chairman, Chief Executive Officer & Product Architect

Great. And that will be roughly a 1000% increase in processing capability compared to the current hardware and so sounds good, a giant of crew despite being a bump, it costs about the same costs volume and car consumption are approximately the same as the current hardware, but it is a tenfold increment in frames per second and improved redundancy as well.

So is it faster, more efficient, or both?

Andrej Karpathy -- Director of AI & Autopilot Vision

Yeah, certainly. Hi, everyone. My name is Andrej Karpathy. I'm the Director of AI here at Tesla. And my team trends all the neural networks that analyze the images streaming in from all the cameras for the Autopilot. For example, these neural networks identify cars, lane lines, traffic signs and so on. The team is incredibly excited about the upcoming upgrade for the Autopilot computer, which Pete briefly talked about. This upgrade allows us to not just run the current neural networks faster, but more importantly, it will allow us to deploy much larger computationally more expensive networks to the fleet.

Frame rate can be limited by the front end data speed of the chip. The amount of processing per frame, less so. v9 is running the equivalent of 5 cameras at full resolution (4 full, 4 1/4) Tesla deploys massive new Autopilot neural net in v9, impressive new capabilities, report says, so only 60% improvement is needed just to get full resolution on all of them. Since they don't need 10x the frame rate, the rest can go to NN depth improvements.

 

[Bladerskb]

Great. And that will be roughly a 1000% increase in processing capability compared to the current hardware and so sounds good.

1000% = 10x increase in processing capability over AP2.
10x of 8-10 = 80-100 TOPS (The actually operations per second might even be way lower than that which is the reason Elon is preventing his team to say it)

/discussion

 

[diplomat33]

Actually it turns out I was mistaken about one of the assumptions I made.

I assumed unconfirmed lane changes REALLY meant unconfirmed lane changes, but it doesn't do that.

You still have to confirm, but all they've really done is move the confirmation from the stalk to the steering wheel.

As I understand it the process works like this.

It does the notification (the beep or vibrate)
Then you confirm the lane change by nudging the steering wheel slightly.
Then it executes the lane change.

Which is different then I was led to believe by the initial description of EAP.

So it's not nearly as scary as I thought it would be.

It feels like Tesla minimized driver input while still keeping NOA as safe as possible. So NOA will still warn the driver before making an auto lane change and give the driver an option to cancel it so it won't take the driver completely by surprise. It won't be the car just making lane changes completely out of the blue which would be dangerous. But at the same time, driver input is kept to a minimum where we basically just need to dismiss the nags like we already do on EAP. And it sounds like with NOA, the car will automate a lot of highway driving now. Basically, the driver will just need to pay attention to the road and dismiss nags from time to time and when the car makes a lane change. The car will handle staying in the lane, accelerating and decelerating with traffic, making lane changes and taking the right exists which is everything you need to get from A to B on the highway. I can't wait to try this new NOA.

 

[diplomat33]

You Tesla fans always coming up with ridiculous theories. 3 years ago it was there is no need for a chip upgrade. Tesla and Elon says this is enough. Elons team is so good they can create efficient software which nvidia cant, meaning they wont need any additional compute. Plus nvidia main goal is to sell more compute so they will exaggerate what's needed.

Fast forward 3 years, it's the same BS but different spin. Now it's we have the best chip but we dont have all the info so it might be 1000tops. Who knows?

You Tesla fans never cease to amuse me.

Dude. I thought you were taking a break from this forum? Why are you so obsessed with trying to point out our "BS"? Just let us actual Tesla owners have our fun and enjoy our cars!

 

[Randy7fx]

It feels like Tesla minimized driver input while still keeping NOA as safe as possible. So NOA will still warn the driver before making an auto lane change and give the driver an option to cancel it so it won't take the driver completely by surprise. It won't be the car just making lane changes completely out of the blue which would be dangerous. But at the same time, driver input is kept to a minimum where we basically just need to dismiss the nags like we already do on EAP. And it sounds like with NOA, the car will automate a lot of highway driving now. Basically, the driver will just need to pay attention to the road and dismiss nags from time to time and when the car makes a lane change. The car will handle staying in the lane, accelerating and decelerating with traffic, making lane changes and taking the right exists which is everything you need to get from A to B on the highway. I can't wait to try this new NOA.

Snippet from the Tesla blog post, which clearly indicates NoA can be set to "enable at start of every trip" which has the option of not requiring a lane change confirmation, contrary to what you indicate above:
*******
In the Autopilot settings menu, a driver can press the Customize Navigate on Autopilot button which will now display three additional settings – Enable at Start of Every Trip, Require Lane Change Confirmation, and Lane Change Notification. Through the Enable at Start of Every Trip setting, Navigate on Autopilot can be set to automatically turn on each time a driver enters a navigation route. Once enabled, anytime a driver is on a highway and uses Autopilot with a location plugged into the navigation bar, the feature will be on by default. If a driver selects ‘No’ to Require Lane Change Confirmation, lane changes will happen automatically, without requiring a driver to confirm them first. Drivers can elect to get notified about an upcoming lane change by receiving an audible chime as well as a default visual prompt. Additionally, all cars made after August 2017 will also have the option to have their steering wheel vibrate for the alert as well.
*******

 

[diplomat33]

Snippet from the Tesla blog post, which clearly indicates NoA can be set to "enable at start of every trip" which has the option of not requiring a lane change confirmation, contrary to what you indicate above:
*******
In the Autopilot settings menu, a driver can press the Customize Navigate on Autopilot button which will now display three additional settings – Enable at Start of Every Trip, Require Lane Change Confirmation, and Lane Change Notification. Through the Enable at Start of Every Trip setting, Navigate on Autopilot can be set to automatically turn on each time a driver enters a navigation route. Once enabled, anytime a driver is on a highway and uses Autopilot with a location plugged into the navigation bar, the feature will be on by default. If a driver selects ‘No’ to Require Lane Change Confirmation, lane changes will happen automatically, without requiring a driver to confirm them first. Drivers can elect to get notified about an upcoming lane change by receiving an audible chime as well as a default visual prompt. Additionally, all cars made after August 2017 will also have the option to have their steering wheel vibrate for the alert as well.
*******

I know that. But the blog also says that you will be required to have your hands on the wheel and a reddit post from an actual user says that the car will require the "apply light force" tug before making an auto lane change. So while auto lane changes will not require driver confirmation, there will still be the nags to make sure you are holding the wheel when it makes the auto lane change.