-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD competition is here
- Thread starter Moderatefan
- Start date
Actually you are using a big part of your brain for vision interpretation. And actually it’s closer to 1:10000 when it comes to human brain vs NVIDIA drive PX2, some would say even more.
Now if you only use 1% of your brain for vision interpretation, that’s still 10 times the power of such a chip.
In 2014 the K Computer, the 4th most powerful computer at that time took 40 minutes to do all the calculations a brain does in one second.
The currently fastest computer could do that in 5 minutes, but that one takes 15 MW to run. So even a 200 kWh Tesla could not run it for one minute.
So let’s assume vision interpretation only takes 1% of your brain. The supercomputer would still take 3 times longer to calculate than a human brain.
So I guess we can agree, that the computer needs every advantage it can get, right?
voip-ninja
Give me some sugar baby
We don't know how much is needed when only using cameras and a radar as the inputs, other than that humans can do it with vision alone, which is why that number was quoted.
Nvidia saying they built a board to do it, when that board has not been demonstrated as being able to do it yet is kind of meaningless.
So you don't need 1 ExaFlop to do FSD. End of discussion, stop spinning about statement you made earlierWe don't know how much is needed when only using cameras and a radar as the inputs, other than that humans can do it with vision alone, which is why that number was quoted.
Nvidia saying they built a board to do it, when that board has not been demonstrated as being able to do it yet is kind of meaningless.
voip-ninja
Give me some sugar baby
Ignorant statement. Driving is largely about predicting behavior and pattern recognition... areas where computers (or as you put it "cars") have typically had zero performance, let alone human level performance.
voip-ninja
Give me some sugar baby
You don't get to decide when the discussion is over, but you can certainly stop participating anytime you like.
Tesla can achieve FSD with the hardware they currently have on the 3?
Because that was the point you and others made about how Tesla doesn't need LIDAR because humans drive with two eyes as inputs.... the problem is you left out the massive processing power of the human brain compared to the hardware that is in the Tesla which is miniscule in comparison.
Maybe Tesla will pull it off... I am highly doubtful and anticipate they will eventually incorporate LIDAR into their sensing suite as costs come down.
The articles I cited above are informative, because they demonstrate the series of departures from Tesla of engineers and executives involved in the self driving program after Musk continually makes promises about what the system can do that the engineers doing the work don't agree with.
Your 1 ExaFlop claim is dead end. You don't use the whole brain to process driving. So saying you need compute power of a brain is false statement.
voip-ninja
Give me some sugar baby
No, it's not a false statement.
What % of your brain is used for driving tasks such as piloting, navigation and most importantly obstacle avoidance, predicting the behavior of other traffic, pedestrians, rabbits, etc?
As someone else who actually did the math pointed out, the Nvidia setup on the Tesla is 1/10,000 the capability of the human brain.
So if you use more than .01% of your brain while driving you still have more compute power than what Tesla is putting into their automobiles to do "full self driving".
If you don't use .01% of your brain while driving then you shouldn't be driving.
I predict that Tesla will continue to sell FSD at $3,000 a whack and continue to install new/better hardware in their cars until some level of FSD is eventually reached 2-3 years down the road.... at that point Tesla can retrofit the hardware into the cars that shipped without it, or refund the money of those who paid for FSD that didn't materialize.
I have 0% confidence that level 4 autonomy is possible with what Tesla has in their cars today. Certainly, the fact that the cameras don't even have the ability to clean themselves (other than the one behind the windshield that can be serviced by the wipers) means it will be extremely inferior to what others are doing with sensors that are much more impervious to road filth, weather, etc.
That’s very true, but that defeats the statement: “A human doesn’t need LIDAR”
A computer is very good at calculating things. So interpreting a LIDAR signal is easier, than interpreting if something is a moving car, or a stationary tree in a split second while moving, with only vision.
Visually, coming over the crest of a hill
a sign over the road is "in the road'.
How can a cam based/visual AI determine
that sign is a big thing, far away,
or a small thing about to
dent your Elon grin?
Map every darn fixed thing, reference the database,
pray for No Random. GeoFenced.
Active would need continuous active recalculation
that relates speed of approach to size of the object...
Yikes.
Things we do automatically with No Thought.
So how do we do that?
Having to think about what we never have to think about
*
Okay, that wasn't really helpful.
Neither is this:
tElon grin?
I have convinced myself this is an inside joke
between Elon and designer dude.
Done perhaps because he really wanted the Roadster
to be called the Lotus Elon
Heh.
Try and unsee that.
a sign over the road is "in the road'.
How can a cam based/visual AI determine
that sign is a big thing, far away,
or a small thing about to
dent your Elon grin?
Map every darn fixed thing, reference the database,
pray for No Random. GeoFenced.
Active would need continuous active recalculation
that relates speed of approach to size of the object...
Yikes.
Things we do automatically with No Thought.
So how do we do that?
Having to think about what we never have to think about
*
Okay, that wasn't really helpful.
Neither is this:
tElon grin?
I have convinced myself this is an inside joke
between Elon and designer dude.
Done perhaps because he really wanted the Roadster
to be called the Lotus Elon
Heh.
Try and unsee that.
Barklikeadog
Active Member
Good posts. Even beyond the Lidar, I think one of the most essential pieces to the puzzle of FSD is having every vehicle on the road communicate with each other. This is far far off in the future.
No. The car has to identify other objects quickly and correctly in any case. Communication would _help_ but can't be relied on.
voip-ninja
Give me some sugar baby
Inter-vehicle communication would be a big plus but the system cannot rely on it as there will be non FSD cars on the roads for decades still and any appropriate advanced system has to account for what those cars are doing.
Thank you very much for underscoring the need for true L4 and L5 driving, and just how dangerous other drivers are!
Oh, and I admit, I've done the same.
insaneoctane
Well-Known Member
I'm not proud and it's a bit scary (and impressive). At least I am using a lot of EAP now....Thank you very much for underscoring the need for true L4 and L5 driving, and just how dangerous other drivers are!
Just curious, has there been any research to see if human vision system acts strictly as a visualization system or do our vision system also acts like a lidar where it can process distances and etc. without first processing the visual information? I am wondering if camera + lidar is really what's needed to duplicate the human vision, rather than just visual only camera.
Our brain doesn't work like radar, or LIDAR. Those need to send out a electromagnetic waves and then they are reflected at an object and travel back to the sensor.
By knowing the time of travel, the computer knows the distance between the sensor and the object and using the doppler effect, how much the wave is compressed, or stretched (think of a fire truck siren passing by, first it sounds faster, then it sounds slower as it passes you by and moves away).
That way radar, or LIDAR can be used to find an object's, location, what direction it's moving and which speed it's moving at.
With vision, you will always need two pictures. Either from two cameras at an angle, or different pictures in time from the same camera. But then you have to find the same object in both picture, first. And that takes a lot of calculating power.
3D maps make it a lot easier, because then the computer can focus on the things that are new to the scene, which makes it a lot easier.
Somehow this is not very encouraging for the upcoming EAP/FSD development
Tesla loses senior Autopilot manager to Lyft’s autonomous driving team
Tesla loses senior Autopilot manager to Lyft’s autonomous driving team
No... but unusual circumstances arise pretty frequently that a person will see as a risk, that a narrow AI won't. What do you do when you see a reckless driver weaving hard back and forth in the rearview mirror? Or a kid staring at their phone between two parked cars, walking toward the road? Or a deer standing on the side of the road? Or a snowplow ahead in the right lane? Or an intersection where the traffic lights aren't working, or where somebody wiped out the stop sign?
A lot of these situations require contextual knowledge about the behavior or people and animals. And that's the Achilles heel of the narrow AI systems. They won't know how to best deal with these sorts of outlier situations. And outlier situations kill people.
I think that eventually these situations will be be dealt with by making modifications to our road system that permit narrow AIs to operate safely. I think we'll see autodrive-only lanes on highways, more restrictions on pedestrian access to major thoroughfares, vehicle to vehicle communications, vehicle to infra-structure communications, highly accurate road surveys that will "certify" a roadway for fully autonomous driving, and legal requirements that autonomous vehicles carry redundant sensor systems.
So... I think we'll get there. But I don't think it will happen just by making the cars smarter. Other things will be required.
Similar threads
