You have a Model 3 which has AP 2.5
The person I was responding to has AP 2.0 which isn't capable of supporting the dash cam feature. It is a SW thing, but it requires HW resources and AP 2.0 lacks the extra SOC that the AP 2.5 hardware has.
-
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
You have a Model 3 which has AP 2.5
The person I was responding to has AP 2.0 which isn't capable of supporting the dash cam feature. It is a SW thing, but it requires HW resources and AP 2.0 lacks the extra SOC that the AP 2.5 hardware has.
Not really, I'm sorry to say. The frame rate is predetermined and fixed, even if the processing power available is higher. The neural network cannot be working at different frame rates. Neither AP2.5, nor AP2 hardware even, is being used to its full capability. Processing power headroom is always required and present in the current design.
I'm wondering if part of their plan is to wait on announcing the new hardware so that they have time to gather miles and train the new NN with fleet data. It's not great if they announce new hardware that is supposed to be amazing and then say "oh, well actually it's not going to be that good for a few months". Reminds me of HW2...
Right, with the exact same neural net, the only difference between HW2 and HW3 is the underlying bugginess of the platform. For example, if the back-up-camera freezes of HW2/2.5 affected more than just the rear camera (HYPOTHETICALLY) or HW3’s Samsung chip has much better image post processing compared to NVIDIA’s SoC, maybe it could make a tiny difference.
But all of that is negligible. What would really start to cause performance differences is a different neural net that takes advantage of HW3’s TRIP capabilities. And that hasn’t happened yet.
SpudLime
Active Member
You guys need to read the other thread or the podcast with the guy who really knows about the NN.
I think your opinion would change.
HW3 and the upgraded NN is vastly different and faster.
I think your opinion would change.
HW3 and the upgraded NN is vastly different and faster.
strangecosmos
Non-Member
There are going to be new neural networks. Andrej Karpathy confirmed this on the Q3 2018 earnings call:
Karpathy’s quote above suggests that all the training is already done. After all, the sensor data captured from cars is uploaded to Tesla’s servers, so it doesn’t matter what computing hardware the cars have. The Hardware 3 NNs can be trained on the exact same dataset as the Hardware 2 NNs. If there is a delay before Autopilot is active on Hardware 3 cars, it will probably because Tesla is just testing that the new hardware and NNs are safe, and not anything to do with training.
At least not with regard to perception. Path planning/driving policy might be a different story. If Tesla is switching over from a hand-coded approach to a neural network approach, and needs lots of human driving data for imitation learning, but needs Hardware 3’s perception networks to get that data with high accuracy (not the raw sensor data but the metadata or mid-level representations), then possibly that could cause a wait before Autopilot is active. But I don’t know why they wouldn’t just ship the old hand-coded software until the new neural network is ready.
“...my team trains all of the neural networks that analyze the images streaming in from all the cameras for 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 [Bannon] 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. The reason this is important is that, it is a common finding in the industry and that we see this as well, is that as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity.
So in other words, we are currently at a place where we trained large neural networks that work very well, but we are not able to deploy them to the fleet due to computational constraints. So, all of this will change with the next iteration of the hardware. And it's a massive step improvement in the compute capability. And the team is incredibly excited to get these networks out there.”
Karpathy’s quote above suggests that all the training is already done. After all, the sensor data captured from cars is uploaded to Tesla’s servers, so it doesn’t matter what computing hardware the cars have. The Hardware 3 NNs can be trained on the exact same dataset as the Hardware 2 NNs. If there is a delay before Autopilot is active on Hardware 3 cars, it will probably because Tesla is just testing that the new hardware and NNs are safe, and not anything to do with training.
At least not with regard to perception. Path planning/driving policy might be a different story. If Tesla is switching over from a hand-coded approach to a neural network approach, and needs lots of human driving data for imitation learning, but needs Hardware 3’s perception networks to get that data with high accuracy (not the raw sensor data but the metadata or mid-level representations), then possibly that could cause a wait before Autopilot is active. But I don’t know why they wouldn’t just ship the old hand-coded software until the new neural network is ready.
Last edited:
We do read those threads. The NN WILL be different, but for FSD, and currently it's NOT different.
If HW3 will improve the NN for Autowipers and sensing rain in the dark I'll be a happy camper 
boonedocks
MS LR Blk/Blk 19”
If this part of the quote is to be believed as accurate and truthful "This upgrade allows us to not just run the current neural networks faster" that is exactly what I have been saying about my HW2.0 car. It seems to handle, braking in particular, EAP much differently than my wife's HW2.5 car. Switching from TACC only to EAP changes it a lot as if its txting or making a grocery list in the background in the way it reacts to stop and go traffic. Driving the two back to back in the same situations, repeatedly, has my informed opinion that HW2.5 is in fact more capable today than HW2.0 in the same situation. It is repeatable not a fluke.
boonedocks
MS LR Blk/Blk 19”
YAY, if it is delayed another 6 months then I will JUST have broken 100,000 miles. It'll be a 100k birthday present for my car.
Hi,
as to the NN accelerator: I here always this device can handle 2000 fps ... WTF ... what does that tell me? It can handle 8 cameras with 250 fps each? Nobody requests that. Ok, I accept that this is more a marketing figure which shows how superior a system is. But this is meaningless. As I said, most important is:
-TOPS per Watt
-latency
-System cost
At that end the NVIDIA solution has the weak point.
And that is TESLA obviously attacking.
On the other hand:
1. 2014 ME AP 1.0
2. 2016 NV AP 2.0
3. 2018 NV AP 2.5
4. 2019 ?? AP 3
That is not continuity or strategy ... poor engineering ... or magicians how show the traditional OEMs what is possible.
Let's see what the AP3 delivers in performance and quality ...
regards,
Frank
as to the NN accelerator: I here always this device can handle 2000 fps ... WTF ... what does that tell me? It can handle 8 cameras with 250 fps each? Nobody requests that. Ok, I accept that this is more a marketing figure which shows how superior a system is. But this is meaningless. As I said, most important is:
-TOPS per Watt
-latency
-System cost
At that end the NVIDIA solution has the weak point.
And that is TESLA obviously attacking.
On the other hand:
1. 2014 ME AP 1.0
2. 2016 NV AP 2.0
3. 2018 NV AP 2.5
4. 2019 ?? AP 3
That is not continuity or strategy ... poor engineering ... or magicians how show the traditional OEMs what is possible.
Let's see what the AP3 delivers in performance and quality ...
regards,
Frank
There is data to indicate the NN is fed two frames at a time. So 2000 fps could be 8 cameras at 125 fps. Or one foot of distance at 85 MPH. Or, if hey did short vs long time base line differential that would be 62 fps.
But yeah, it depends how much processing you do on that frame.
No. AP 2 came out in October 2016. AP 2.5 came out in August 2017.
Also, AP1 was 1 camera. AP2+ all have 8 cameras so far. There was online mention of only "using" 4 cameras but the cars have all have 8 cameras since 10/2016 and they use all cameras for EAP.
Last edited:
Is AP2.5 the same as what the Model 3 comes with? I never had an S or X but I seem to recall reading one of them, AP1 or maybe 2 (as well) used the autopilot system from either NVidia or Google (Waymo). Consequently all of the millions of drive miles accumulated were incompatible with that from the AP 2.5 meaning Tesla's base of acquired training had to start over from scratch. Does that mean that version of the S and X are orphaned as the entire technology is different?
The original AP1 system is a single camera with a dedicated Mobileye chip ruining a mix of Mobileye and Tesla code.
Mobileye wanted a bunch more money per chip for the next contract and access to Tesla's data and development effort, so Tesla went their own way.
Everything since then has been Tesla software, mostly running on NVidia hardware so far - which as I understand it is general ARM rather than something unique that might be hard to port code to new hardware (the Mobileye chip is unique hardware and I'm not sure what if anything Tesla was able to port out of their part of the IP.)
I'm excited too, just afraid that retrofits would be on the backburner for quite some time
That's the keyword I was trying to remember, MobileEye. I was trying to find where I had read those earlier miles could not be counted on the new hardware, not sure how new as in 2.5 or 2.0). I can see the graph in my mind's eye and it was explaining how the older data was not transferable to the newer technology. Of course now, with the M3 fleet there are many more sources of training models. I don't recall what the delta is between the pre-mid 2018 tesla fleet and the aggregate M3 fleet but it's a lot more miles. My understanding too is that even if an owner didn't order EAP, the car is collecting training miles as it still has the cameras and sensors.
Tesla Autopilot is not yet approved for Model 3 in Europe – but don’t panic
So at this point they are officially saying it is the car (model 3) that hasn’t passed regulatory approval for autopilot. Not new autopilot hardware.
Unless Tesla are trying to go full blown stealth mode, it sounds like all cars still have hw2*. Are there new X and S cars that have autopilot working arriving in Europe st the same time?
So at this point they are officially saying it is the car (model 3) that hasn’t passed regulatory approval for autopilot. Not new autopilot hardware.
Unless Tesla are trying to go full blown stealth mode, it sounds like all cars still have hw2*. Are there new X and S cars that have autopilot working arriving in Europe st the same time?
Similar threads
- Article
