Slashdot Mirror
Driverless Cars Are Giving Engineers a Fuel Economy Headache (bloomberg.com)
schwit1 shares a report from Bloomberg: Judging from General Motors' test cars and Elon Musk's predictions, the world is headed toward a future that's both driverless and all-electric. In reality, autonomy and battery power could end up being at odds. That's because self-driving technology is a huge power drain. Some of today's prototypes for fully autonomous systems consume two to four kilowatts of electricity -- the equivalent of having 50 to 100 laptops continuously running in the trunk, according to BorgWarner Inc. The supplier of vehicle propulsion systems expects the first autonomous cars -- likely robotaxis that are constantly on the road -- will be too energy-hungry to run on battery power alone. A fully autonomous subcompact car like a Honda Fit, for example, will get 54.6 miles to the gallon in 2025 in the best-case scenario, more than 5 miles below the U.S. emissions target, according to BorgWarner. A small pickup or SUV would be at 45.8 mpg, versus a target of 50. Engineers don't have much time to resolve this, as companies are planning to deploy their first fully self-driving cars in the next couple of years. One way for automakers to meet the power-hungry needs of self-driving systems will be to use gasoline-electric hybrid models rather than purely electric cars, said Mary Gustanski, chief technology officer of supplier Delphi Automotive Plc's powertrain business.
Really? These engineers didn't consider that processing power is constantly shrinking and becoming more efficient?
I feel like I'm reading an article from 30 years ago about how computers will never fit inside your home because the take up large rooms!
My PC uses roughly 400W and can handle over 100 self-driving cars in GTA easily.
The answer is simple: Outsource the processing power to an overseas call center where your virtual driver (Let's call him 'Steve') will steer the car in a simulator-like environment. A few webcams around the car will certainly use less power than that huge LIDAR pod on top of WAYMO cars.
I run a small cluster on less than 2kW of power which would include 3 nodes with each 16 cores, 256GB RAM, 1TB in SSDs and a Tesla GPU.
Just to give you an idea, 2kW of power on your standard 12V car requires a current of 167A.
I'm not even sure what you would have to put in to get a package of 2kW, that's 2000W. Camera's take up less than 2W each, even if you need 10 of them, you're still at only 20W LIDAR take up perhaps 10W each, one for each end of the car is another 20W. Give or take another 10W for various sensors, you've got 50W. A computer unit with a decent GPU for highly parallel tasks can be done in 200W.
Custom electronics and digital signage for your business: www.evcircuits.com
Self driving cars in Europe might be neat where they don't have transportation shortages. In America, our mass-transportation infrastructure is non-existent. Except for a handful of cities, you have to own a car in order to simply function in society, or you have to find a job that lets you work from home or live in a very limited area of town.
I wrote a post about this a while back:
http://penguindreams.org/blog/self-driving-cars-will-not-solve-the-transportation-problem/
Basically even if you had Interstates which only allowed self-driving cars and all of them could travel at over 120kph bumper-to-bumper and all of them were filled with four people each, you still wouldn't even get to 10% of the capacity of a traditional rail system, running on a single track, with trains arriving at 5 minute intervals (and most cities with rail systems have them arriving at 2 min intervals during rush hour. London has several automated trains. Singapore is fully automated).
Before we start dumping billions into subsidizing self driving cars, how about we build up our self-driving train tech; a known technology which currently exists and transports millions of people every day.
1) Despite the general belief, driverless has NOTHING to do with all electric. Yes, there will be a competition.
2) Eventually, the computing power and sensor tech will improve power efficiency, allowing both in the same car. Expect that to happen 5-10 years after the first one becomes ubiquitous.
3) I bet they are underestimating how much power they will save by getting humans out of the loop.
4) You can probably decided to turn the AI off to drive long distances, thereby saving battery life.
excitingthingstodo.blogspot.com
to crowdsource driverless cars. Obligatory xkcd link.
https://xkcd.com/1897/
... the equivalent of having 50 to 100 laptops continuously running in the trunk, according to BorgWarner Inc.
Apparently, Resistance isn't futile it's V / I.
It must have been something you assimilated. . . .
Consider: Many long-haul trucks operate mostly on well maintained highways between distribution centers. Ten or more of them could be linked together scanning for issues, and communicating with each other, possibly with a lead car in the front that can react to accidents and incidents before they even pass, similar to over-sized loads on the road today. Even if not driving between distro centers, they can use existing rest-areas as stopping points where local drivers can take over in shifts.
and then network costs / power will go up
So they're saying we should be wary of the consumptive nature of the hive mind in the trunk?
I should have expected no less from Borg Warner.
Two to four kilowatts in an electric automotive application? So what's the problem? Accelerating a ton of metal takes a HYSTERICAL amount of electric power.
A horsepower is almost exactly 3/4 of a kilowatt. So four kilowatts is 5 1/3 horsepower.
A vehicle just cruising at highway speeds takes teens of HP just to fight wind resistance and rolling friction. (Numbers from a long while back, so aerodynamic improvements may have brought it down a bit, though I doubt it's under ten HP.) Accelerating takes a lot more, like over 100 HP if you're not going to hold up traffic. Then there's hill and mountain climbing.
Meanwhile the motors and their controllers are not 100% efficient at converting electricity to motion, but the computer load is electric.
So start with the computers adding another 30% to the load for the long-term cruise - about the minimum cycle for the car. Yes, that will cost nontrivial miles per kilowatt hour. But the computer algorithms might save more than that by driving more efficiently than a random human.
Now look at those computers. They're laboratory experiments. Getting the car to go where you want and not kill or break something on the way is the hard engineering task right now. So they're not optimized for power consumption. They're designed to have enough processing power that they let the engineers try out any stack of new bright ideas. Once the algorithms are debugged and a desirable set picked, so the engineers know the peak computational load, you can be sure there will be a stage where they design a computer that can handle the crisis-peak load, but not a WHOLE lot more, and can scale back its power consumption when it doesn't have to think that hard.
Meanwhile, Moore's law isn't dead yet, so the load for a given amount of crunch is still dropping.
I'm betting that, by the time the tech is ready to be deployed beyond the rich early-adopter stage, the fuel energy cost for the computers will be substantially less that the fuel energy cost of carrying around the extra weight of a chauffeur skilled enough to drive as well as his silicon competition.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
This is calculations based on wild assumptions about what is actually required for self-driving. If Tesla is right, then they can do it with the hardware that is already shipping, so power is not an issue--they just need to finish the software. You can also look at what Nvidia just released that they're billing as being designed for self-driving cars. AMD is apparently working on a similar product.
This sounds like a typical naysayer making stuff up to get attention (and advertising hits).
Sure, you can hack my toaster and burn my crumpet all you like.
No fucking way am I getting into a computer-controlled car...
What kind of a person gets excited about a driverless car? What's the fun in that? I just don't get it.
The NVidia or other GPU approach to AI is too flexible for this application. It needs a more purpose-built chip. Perhaps something like IBM's TrueNorth or even a mixed analog/digital NN approach.
NNs in general have potential to be much more power efficient than traditional computing with vectors makes it appear, but not when we use traditional computing techniques to simulate the NN.
We will see this evolve quickly as the market appears. It's still quick and dirty time right now.
A Tesla Model S has a 90 kilowatt hour battery life for it's 275 mile range. A kilowatt hour is a thousand watts, and let's assume you're going 60mph. For argument, and maths, sake we'll assume you're going above efficiency and say you'll get 240 miles out of that. That's 4 hours, so 22.5 kilowatts an hour. That's a powerdraw of 22 thousand watts in 1 hour. The new self driving chip announced by Nvidia only draws 500 watts, that's 500 watts in an hour. Or better yet, here's the empirical evidence of Tesla owners discussing their average watts/min usage: https://forums.tesla.com/forum...
Even there with more efficiency, the new Nvidia chip uses in an hour less energy than the car itself uses in 2 minutes. This article is absolute bullshit, they had 1 damned thing their job required and they didn't do it. Self driving electric cars are perfectly mathematically sounds.
Yes, Self-Flying Fuel Cell Electric Cars are only a few years away.
"His name was James Damore."
Instead of worrying about the efficiency of something that might be made in 2025, let's worry about how much power these things consume when someone actually builds a fully functional one that is for sale.
Then we will at least have a metric for comparison and competition.
Keep in that current electric cars manage to drive 100km with around 15kWh, needing another 4 kWh for the autonomous hardware will decrease the already pretty pathetic range by quite a bit.
Your petrol car may have 100hp, but unless you're going down the highway full throttle, you're using much less.
This sounds stupid, and also, even if it were true then this is something that gets solved the more autonomous cars you get on the road:
* You get more computer driven cars on the road, the roads will become more computer-car friendly. Think of signs or markings on the road that will tell the car where it is, or what kind of hazards are coming up via for example a QR code.
* Once you get more driver-less cars you get inter-car communication. Car up ahead does the full sensor sweeps and pass the results down the train, saving power for all the other cars.
* Once most cars are driverless, you'll get dedicated driverless car lanes which should be made easy to navigate for them
* And when human driven cars are totally extinct I expect cars to be able to navigate roads with something equal to a raspberry pi today.
obviously they are running alpha versions of the hard and software, yes, it will be 30 laptops in the boot. the production model will be nothing like this.
just look at alpha versions of computers from the old days, it filled several tables of boards and wires everywhere, still in the end you get a nice small box.
On a long enough timeline, the survival rate for everyone drops to zero.
I thought there were efficiencies to be had from a self-driving car being able to optimize acceleration, braking, and with car-car communication, greatly reducing a lot of the stop/go in congestion. It should be able to reduce the power consumption of the engine/motors over the way humans accelerate, brake, and speed up and slow down too often.
These won't make up all the losses to self drive, but won't they offset some meaningful percentage? 5% or so power consumption from the drive system through efficient power application should provide a meaningful buffer for the self-drive system.
I also wonder if the self-drive system couldn't partially sleep some part of the sensor environment and processing through some complex decision system involving time of day, real-time congestion reports, road condition, weather, road geography and car-car reporting data. Basically if you're on a long, straight street with no detectable traffic at 3 am in clear weather, couldn't the car dial back the sensor operation to draw less power? Or even if it kept them all running, obtain some efficiency because there's just less change data?
Making a Tesla autonomous is not going to require something in the range of 10 times as much computing power than what 'auto pilot' currently needs.
Depends on whether this additionnal processing capability will be fighting against diminishing returns.
Might by that the last 10% of processing capability to get it right require 90% more computations.
On the other hand, the processes used (deep-neural nets, etc) tend to be extremely friendly to parallel processing (they love multicore, GPUs, etc.) and to specific ASICs (think GPUs or even Google's TPU). They aren't as much reliant on huge technological progress (you don't necessarily need a single chip clocked at 50Ghz and built with a 0.1 nm process).
So even if today's prototype consume as much as a small beowulf cluster of linux nodes in the trunk, chances are by the time the thing goes into production, it could be handled by much reasonnably power-hungry hardware.
(Also, none of the current fleets of autonomous car deployed in production is suffering due to power budget.)
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
It seems a bit disingenuous to suggest that *experimental* systems that are running on general purpose computing platforms are going to be the actual endgame and then using that to decide that self driving cars are going to have a problem with power consumption. Of course, I don't know what the specific tech involved in each self drive system actually is. However, because everything is experimental at this point, it is pretty much certain that nothing is optimized. It would be a waste of resources to optimize at this stage of the game. I can guarantee that the makers of self driving cars are going to be doing everything they can to minimize the power draw of their control systems simply because they need to compete on *range*. Anything that drains the battery is on the table for optimization or removal if they can improve range.
Once they work out exactly what sensors they need, how to process the data, and all that, then no doubt most of that processing smarts will move down into custom devices optimized for the specific tasks which will almost certainly remove a lot of the power consumption. That's not to say there won't be more power consumption in a self driving car than a human operated one. After all, even the human brain takes a nontrivial amount of energy to operate!
If it works in theory, try something else in practice.
now that is cool...
https://www.youtube.com/watch?...
At the very least judicious hypermiling by AI will save a lot of energy
love is just extroverted narcissism