Right-O. Plus 300-400 W for gpu gives less than 500 W. Go team.
Right-O. Plus 300-400 W for gpu gives less than 500 W. Go team.
It absolutely does.Did not you want to write "Less than 50W?"Autopilot is not heating yet...
Perform this experiment: open the car door, leave the fob inside and walk away for 15 minutes. Then return, open the glove box and see how warm it is inside.
If you take the ape fans off and power it, even in absolute idle the whole housing heats to ~60C in like 10 minutes.
rooter
Member
The -charging- voltage is about 14.2. The -battery- voltage is about 12.5.
You must charge at a higher voltage than the battery or, eh, it won't charge.
One learning here is in round numbers the relative power cost of the AP system to the total load going down the road.
I noticed yesterday that I was doing 350 wh/mile at 60 mph with AP, a bit cabin heat, etc. That’s 21 kWh used in an hour. During that hour, assuming an AP load of 500 w, AP used 0.5 kWh. That put AP at around a quarter of
I noticed yesterday that I was doing 350 wh/mile at 60 mph with AP, a bit cabin heat, etc. That’s 21 kWh used in an hour. During that hour, assuming an AP load of 500 w, AP used 0.5 kWh. That put AP at around a quarter of
Oops. Final result was supposed to be AP running at 2.5 percent of total load. Hit wrong button : )
Hang on, let's just go back to the published nVidia numbers. Don't they claim those chips can consume as much at 400w when running at full load? There's no way that thing is processing what it is (or processing FSD) at 40watts. We are talking about something more powerful than a 10x0 series graphics card here.
The chip they have is an equivalent to 1060 card pretty much.
~40 watts is idle consumption of the whole unit (including FANs, Linux two SoCs and the rtos one, ethernet switch, deserializers, RAM and all that stuff)
Tesla NEVER had the ~500W Nvidia machine. People extrapolated this based on unbounded optimism, because on the other hand you had the possibility that Tesla was shipping a chip that wasn't capable of FSD by Nvidia's standards. Which is true.
Yes, the charger puts out more than the resting battery voltage to charge the battery. However, you are implying a 1.7 volt drop on the wires connecting the charger to the battery.
When the car is on, everything 12V is being supplied by the DC-DC converter/ charger. The charger could determine the 12V battery is full and then lower the system voltage, or just run 14.2V all the time. In either case, the voltage at the battery terminals will be very close to the charger output (I*R drop) with the additional battery cell voltage a result of the internal resistance and electrochemical reactions of charging.
System voltage = charger voltage when charging.
System voltage = battery voltage when not charging.
Edit/ addendum: The power used by the AP/ MCU will stay fairly constant over voltage due to the use of DC-DC power supplies.
Well that's interesting. And disappointing. And basically proves my unpopular claim from the beginning that Tesla released AP2.0 hardware knowing it wouldn't be enough for FSD. The next shoe to drop will be that the sensor suite isn't up to the task in all common weather conditions.
I've been saying this since day 1 as well, but no one seems to listen or care. The cameras are basically useless in even minor rain/fog/etc. If they would have included, at a minimum, the rear corner radars, then I would have given them the benefit of the doubt. There's absolutely no way the currently hardware+sensor suite can get close to FSD without absolutely perfect conditions.
But seriously, just try to use your backup camera when it gets just a little wet. Then cover all of your windows and try to drive using just that. God speed.
The dynamic range on the cameras is impressive, but it's still not good enough to see many vehicles approaching on the sides/behind when it's dark outside. It's easily fooled by lights in the distance and other things and lots of the time won't detect a vehicle until it's basically within reach-out-the-window-and-touch-it distance. Radars would have solved that, since they'd have a reference to go with for the limited visual data.
But nothing will make this hardware/sensor suite work well in the rain. It can do autosteer ok in the rain right now because of the forward radar augmenting the forward vision data. Beyond that, FSD isn't happening in the rain/snow/whatever. So, you'll set a course in your FSD Tesla when it's sunny, 5 miles down the road an unexpected Florida sun-shower pops up, and the car will be screwed.
I guess Tesla still hasn't figured out that not everyone lives in California. Pretty obvious in my Model 3 with the latest firmware where today I had to just constantly keep hitting the wiper-swipe button for intermittent wipers, despite being set to AUTO the whole time. The windshield basically has to be covered in rain completely to the point where you can't make out a car directly ahead before the auto wipers come on.
Lost a lot of faith in Tesla on this stuff over the years. Over sell, under deliver. All of the time with minimal exceptions.
Rear radar would have been acceptable to address a large majority of driving conditions and improve existing performance. But their hubris prevented them (EM rather). But this truly is a car designed for California. Just look at the model 3 windows and door handles sticking in ice.....
JRP3
Hyperactive Member
I've had sticking door handles and windows in ice on other vehicles so that's not a very good example.
pupik
Member
a real example from my diagnostic software with SQL recording...
...standard consumption from 12V circuit. Model S 2017 with ECU1 and AP2 !
...standard consumption from 12V circuit. Model S 2017 with ECU1 and AP2 !
robby
Member
Do we know yet how an optimized neural net will process those rainy/foggy images compared to a human? I wouldn't rule out that there is some signal in what looks like noise to us. Nvidia's '16 DRIVENet Demo was pretty impressive in this regard[1], and obviously these things only get better with more data.
Certainly AP2-3 won't work in all conditions, but I think it's too early to say where the point is at which (current sensors + optimized software) will be statistically less safe than a human driver. That's based on what I know at least, but maybe not on what you know.
[1] See 7:20
Yeah, the "neural network" will be able to decide in a fraction of the time a human could, that it can't see worth a damn.
