Youtube: Telsa Autonomy Day - Tesla custom AI chip for self driving

S

Snowdog

[H]F Junkie
Joined
Apr 22, 2006
Messages
11,262
A few years ago, when Tesla said it was going to develop their own chip for the AI to to do self driving on their car, this was met with some skepticism. They reveal a lot of detail here:

6 Billion Transistors: 260mm2 - 14nm
72 TOPS from the AI engine
12 ARM A72 Cores
600 GFlop GPU (it's all they need. It's not a game machine)

It's a pretty dry presentation:



If you want to skip ahead to die section breakdown it starts at 1:20:22
 
Watched this. Was legitimately impressed in that "..............................................well ok then, damn." way. Nicely done Musk, you GO Frank Scorpio!
 
Spartacus09 said:
What was the performance of the previous chip for the self driving?
Click to expand...

The big job here is visual recognition, and how many frames per second they can process.

On the old NVidia Drive HW, they could process 110 FPS, while on the new HW, they can process ~2300 FPS.
They mention this around 1:32:07

Or 21 TerraOps NVidia vs ~140 TerraOps in house.

Though the new HW has redundant processors and it is counting both of them, when in reality they would merely be duplicating each others work. I don't believe NVidia Drive network has redundant processors.

Even if you cut that in half, it's still ~10x faster and visual recognition than the old drive computer.
 
Last edited:
Snowdog said:
The big job here is visual recognition, and how many frames per second they can process.

On the old NVidia Drive HW, they could process 110 FPS, while on the new HW, they can process ~2300 FPS.
They mention this around 1:32:07

Or 21 TerraOps NVidia vs ~140 TerraOps in house.

Though the new HW has redundant processors and it is counting both of them, when in reality they would merely be duplicating each others work. I don't believe NVidia Drive network has redundant processors.

Even if you cut that in half, it's still ~10x faster and visual recognition than the old drive computer.
Click to expand...
Thanks for the info wasn't knocking it just wondering how the specs from their original chip to their task specific chip compared.
 
I call bullshit on self driving cars. Maybe where Elon lives, in sunny California, scientists and engineers can largely automate driving and come up with a safe final product. However, I'm very skeptical, and where I live it is downright impossible. There is no way in hell any self driving car is going to function when it can't tell where the road is under inches of snow. See this example of what it is like to drive in a tiny bit of snow. It gets much, much more difficult when there are several inches on the road and it hasn't been plowed or treated yet.


In the snow, a good driver has to carefully monitor their speed because traction can be highly variable and stopping distances are greatly increased. Turning or even changing lanes is tricky. Most of the time you can't see lane markers, so roads that typically have 3 lanes each direction might be down to 2 and they aren't exactly well defined or straight.

The reality is we're going about this all backwards. It is like trying to come up with trains that can drive on the road. Nobody is dumb enough to attempt that. Instead of designing cars to autonomously drive on existing roads, we need to focus on a system of roads that are deliberately built for self driving cars. Until that happens, this is all a literal pipe dream (light one up for me, Elon).
 
doublejack said:
I call bullshit on self driving cars. Maybe where Elon lives, in sunny California, scientists and engineers can largely automate driving and come up with a safe final product. However, I'm very skeptical, and where I live it is downright impossible. There is no way in hell any self driving car is going to function when it can't tell where the road is under inches of snow. See this example of what it is like to drive in a tiny bit of snow. It gets much, much more difficult when there are several inches on the road and it hasn't been plowed or treated yet.


In the snow, a good driver has to carefully monitor their speed because traction can be highly variable and stopping distances are greatly increased. Turning or even changing lanes is tricky. Most of the time you can't see lane markers, so roads that typically have 3 lanes each direction might be down to 2 and they aren't exactly well defined or straight.

The reality is we're going about this all backwards. It is like trying to come up with trains that can drive on the road. Nobody is dumb enough to attempt that. Instead of designing cars to autonomously drive on existing roads, we need to focus on a system of roads that are deliberately built for self driving cars. Until that happens, this is all a literal pipe dream (light one up for me, Elon).
Click to expand...


To be fair the electric cars in general aren't currently well suited for the cold especially with the capacity hit from cold weather driving.
 
doublejack said:
I call bullshit on self driving cars. Maybe where Elon lives, in sunny California, scientists and engineers can largely automate driving and come up with a safe final product. However, I'm very skeptical, and where I live it is downright impossible. There is no way in hell any self driving car is going to function when it can't tell where the road is under inches of snow. See this example of what it is like to drive in a tiny bit of snow. It gets much, much more difficult when there are several inches on the road and it hasn't been plowed or treated yet.


In the snow, a good driver has to carefully monitor their speed because traction can be highly variable and stopping distances are greatly increased. Turning or even changing lanes is tricky. Most of the time you can't see lane markers, so roads that typically have 3 lanes each direction might be down to 2 and they aren't exactly well defined or straight.

The reality is we're going about this all backwards. It is like trying to come up with trains that can drive on the road. Nobody is dumb enough to attempt that. Instead of designing cars to autonomously drive on existing roads, we need to focus on a system of roads that are deliberately built for self driving cars. Until that happens, this is all a literal pipe dream (light one up for me, Elon).
Click to expand...


I think you fail to grasp the implication of Neural Network based machine learning. Why makes you think a human will always be better at detecting where the road is, or be better at managing poor traction in these kinds of snowy conditions? These are solvable problems with machine learning.

The snow question was asked after presentation 2. Telsa says that so far, they haven't done specific snow training, but they can still do an ok job in snow, because they not totally dependent on lane markers. They are more dependent on "driveable space" which takes in a variety of factors, not just lane markers. Obvious more specific snow training will improve the snow performance.
 
Snowdog said:
I think you fail to grasp the implication of Neural Network based machine learning. Why makes you think a human will always be better at detecting where the road is, or be better at managing poor traction in these kinds of snowy conditions? These are solvable problems with machine learning.

The snow question was asked after presentation 2. Telsa says that so far, they haven't done specific snow training, but they can still do an ok job in snow, because they not totally dependent on lane markers. They are more dependent on "driveable space" which takes in a variety of factors, not just lane markers. Obvious more specific snow training will improve the snow performance.
Click to expand...

I'm not the one failing to grasp the challenge here. Humans improvise. That's why a 3 lane roadway may have 2 usable lanes that meander when there's a lot of snow. Someone with four wheel drive lays down a set of tracks making the road somewhat passable, and everyone just kind of follows the trail. Nobody is going to come up with a neural network that's going to replicate that. Instead, we're going to get spinouts everywhere as vehicles careen out of control because the AI is trying to force vehicles to drive where it is impassable. What's the machine learning solution to that? Shut down all the cars when several crash? That'll be super popular.

When it comes to taking multiple sensory inputs and making quick decisions in a complex environment, I'll take the human brain every single time over a machine. There's no question. Robotics has been a field of study for decades now and very simple tasks, like picking up and handling soft objects, is still not perfected. Maybe someday AI will be as good as a human brain when it comes to driving decisions, but it won't happen in my lifetime. Hell, Boeing and the 737 Max crashes is perfect example of a far less complex scenario that engineers have totally screwed the pooch on. Give me a pilot and unplug the autopilot, please. If we haven't gotten flying right then there's no chance driving is going to happen.

Another point you're glossing over or missing is that your neural network can be used to make human drivers even better. Every heard of or use waze? That's a start. This is not an either-or scenario. A human with an AI navigator gives the best of both worlds.

Regarding Tesla and their statements on snow... they are full of shit. Laughably so. They haven't done any specific snow training? GTFO. Their system will fail, period. I have seen so many competent drivers, several with years of experience in warmer climates, utterly fail in their first winter. Tesla's system will be no different.

Again, the solution here is to build purpose built roads for self driving cars. Going at it bass ackwards is only going to keep killing people.
 
doublejack said:
I'm not the one failing to grasp the challenge here. Humans improvise. That's why a 3 lane roadway may have 2 usable lanes that meander when there's a lot of snow. Someone with four wheel drive lays down a set of tracks making the road somewhat passable, and everyone just kind of follows the trail. Nobody is going to come up with a neural network that's going to replicate that. Instead, we're going to get spinouts everywhere as vehicles careen out of control because the AI is trying to force vehicles to drive where it is impassable. What's the machine learning solution to that? Shut down all the cars when several crash? That'll be super popular.

When it comes to taking multiple sensory inputs and making quick decisions in a complex environment, I'll take the human brain every single time over a machine. There's no question. Robotics has been a field of study for decades now and very simple tasks, like picking up and handling soft objects, is still not perfected. Maybe someday AI will be as good as a human brain when it comes to driving decisions, but it won't happen in my lifetime. Hell, Boeing and the 737 Max crashes is perfect example of a far less complex scenario that engineers have totally screwed the pooch on. Give me a pilot and unplug the autopilot, please. If we haven't gotten flying right then there's no chance driving is going to happen.

Another point you're glossing over or missing is that your neural network can be used to make human drivers even better. Every heard of or use waze? That's a start. This is not an either-or scenario. A human with an AI navigator gives the best of both worlds.

Regarding Tesla and their statements on snow... they are full of shit. Laughably so. They haven't done any specific snow training? GTFO. Their system will fail, period. I have seen so many competent drivers, several with years of experience in warmer climates, utterly fail in their first winter. Tesla's system will be no different.

Again, the solution here is to build purpose built roads for self driving cars. Going at it bass ackwards is only going to keep killing people.
Click to expand...

Hopelessly stuck in the past. A computer will be able to monitor and react with 360 degree vision a hundred times/second. While a human struggles reacting once/second to something happening in the only direction he is looking.
 
You must log in or register to reply here.
Back
Top