The electricity bill could be insane
I covered this some in-
Let’s look into the poker AI which won this year. Pinning the computer power down has been oddly difficult. Bloomberg said 3,330 high end MacBooks. This link says 10,000 computers. Both are reliable sources. Carnegie Mellon says the supercomputer used 600 of its 752 regular memory nodes. Each at 128 GB of RAM.
The Carnegie Mellon source qualifies a high end laptop as 16 GB of memory. (Not storage. That’s different.) For those who don’t know, this is a seriously high end laptop. For comparison, my brand new MacBook was ~$1300, is only 8 GB, and the thing flies. A 16 GB Apple MacBook is two grand.
All the various numbers are not to confuse anyone. It’s meant to show: this is a ton of computer power.
For 600 of the 752 regular memory nodes, as we said each node had 128 GB of memory, and 2 Intel Haswell (E5-2695 v3) chips. One of these chips uses 120 watts.
- 600 nodes * 2 chips per node * 120 watts per chip = 144,000 watts
This is unlikely all the electricity the supercomputer used. To avoid going off the reservation we’ll stop here, but you can see more details on power consumption of high end computers below-
–Taming the energy use of gaming computers
And how much does the human brain use?
- 20 watts
When I see these statements about AI being better than humans at an activity, sure, technically it’s true. But it’s like saying 144,000 watts / 20 watts per human = 7,200 human brains are better than one. O rly??
The poker AI kept working at night, after the daily games were completed. We could think of it as practicing for the next day. It was running 24/7.
Converting watts to kilowatt-hours,
- 144,000 = 3,456 kilowatt-hours
In California -where AI most prominent- electricity costs 15 cents per kilowatt-hour.
- 3,456 * 0.15 = $518
That’s for one day of work. Call it $135,000 per year. While one can make an argument the AI might not need all this power to beat the players, the other argument is, again, we have an AI which could not pick up cards, could not play within a group, could not quickly play other versions of poker, and had to play at what was described as an agonizingly slow pace.
We also haven’t included the $10 million it took to build the supercomputer. The electricity it took to initially train the program. Never mind the human labor- five years worth! Giving us a number that will be much larger than what a professional poker player spends training themselves. If we project out 10 years, even ignoring the training of the system, where we say each subsequent system already has access to the training,
- ($135,000 electricity per year * 10 years) + $10 million to build the system = $11,350,000
Few employees are going to have that type of earnings over 10 years, negating the labor argument: if the AI is cheaper than human labor, it’s much easier to replace the human.
If we take the assumption personal training / handling clients / [insert other professions] are as complicated as playing poker, we have a system which is going to easily be more expensive than most of those professionals.
Even if the system is better, how much better is it? Is that better worth e.g. 100, 200, x % more in cost?
That said, if we reduce the cost of building the system in half, and make the system able to do two or three jobs instead of one -while the system played slow, it did play four separate people / games at a time- we quickly get into a pragmatic investment. High earners beware.
On the other side of the spectrum, Tesla announced their cars now have full self-driving hardware capabilities. The NVIDIA computers they use come in at $15,000. That’s just the computer. Not the radar, cameras, other sensors, software.
-> I’m pretty shocked Tesla is going to be able to get cars as cheap as they say they are, despite adding all this hardware. We’re talking like a third of the price of the Model 3 being self-driving hardware. It looks like a customer would get the hardware, but still have to pay to activate the software. A payment which could be upwards of $10,000. Either way, if they can do it hardware wise, kudos to them.
But an investment much less than the poker playing machine. If we’re talking truly replacing humans with jobs involving handling clients / patients, jobs involving more intricate movements than driving a car, the cost is well above Tesla’s car. (Especially because we haven’t factored in language processing.) Assuming the car lasts seven years with no major work needed, e.g. new battery, that’s minimum $35,000 / 7 = $5,000 a year.
The range of the Model 3, their cheapest car, is projected at 200 miles. That’s ideal driving conditions. Let’s say 180 miles is the common range. Tesla estimates 180 miles of charging => $7.22
- $7.22 / 180 = $0.04 per mile
Driving 70 miles per hour,
- 180 miles / 70 miles per hour = 2.6 hours
That’s about three full charges per eight hour day. $22.
-> A big drawback here is if we’re in a gym or hospital, we either would need multiple AIs or an always plugged in one. Otherwise we’d have big gaps due to waiting on charging. To charge the 180 miles worth of range would require a minimum three hours. We would actually need two cars to fill a contiguous 8 hour day of work.
- $22 per workday * 50 weeks per work year * 5 days per work week = $5,500
Plus our $5,000 per car per year and we’re at $10,500 per year. Even if we add the $10,000 autonomous activation fee (amortized over seven years), even if we double this by adding a second car, we’re low on the salary ladder.
Where is personal training, or a given profession, on the spectrum? Closer to poker players or driving? Some ways to look at this:
- Any poker player could drive people around for a living. Not any driver could play poker for a living.
- You can drive while doing a lot of other stuff with your brain. For instance, being on the phone.
- Billions of people can drive. Billions of people can’t be professional poker players.
Where if you can do it with half your brain turned off / on other things, or if billions of people could do the job, then perhaps it’s closer to driving.
What’s all but certain: if you drive a car for a living…things aren’t looking good. There is a significant financial incentive to get rid of you. At least in most cities in the U.S. California is not only ripe for the industry due to culture, but also due to higher salaries.
Tangent- the power / heat argument
The driving numbers above, presumably, don’t include full autonomy. Will the battery last as long during full autonomy? The Nvidia computer powering the self-driving capability is equivalent to 150 MacBook Pros, yet only listed at 250 Watts. That’s barely more than a watt per MacBook Pro. (Nvidia’s focuses on GPUs, which are more power efficient than CPUs.)
What we can say so far is the electricity wouldn’t be cost prohibitive, nor battery draining, but it is still more than 10x the power of a brain.
We know the brain only needs 20 watts. Granted, a person is more than a brain. One relevant detail here is whenever we convert energy, whenever we use machines to do something for us, we release heat. While the focus on the planet warming is currently on CO2 emissions, if we keep going the way we are, down the line how much power we produce period is going to be a concern. Produce power => produce heat.
Say a human uses the proverbial two thousand kilocalories per day.
- 2000 kcal / 24 hours = 83 kcal per hour
That converts to 96 watts. If we use a self-driving car, we need 250 watts. Suddenly every car which is autopilot rather than human driven needs that wattage difference. When you’re talking billions of cars, that might matter!
-> That’s something like 2% more watts per electric car. There is enough math in this post already. Just want to get the idea across that while say, an electric car may be more efficient than a gasoline, when we strictly look at heat production, that efficiency is going to get cut into with autonomous driving, due to coinciding with the emerging electric car market.
“But autopilot will make driving more efficient. Less wrong turns, less stopping and starting.” There’s no guarantee of that. Some have argued autopilot will allow cars to have a much higher speed limit, which would increase power demands.
In fact, this is a friendly estimate. Never mind not including the intensified data center demands from the machine learning responsible for autopilot ability,
-> Google employee predicts talking to phone for three minutes a day can triple their data center demands. That could quickly bring us to ~2% of our energy being for data centers, to 6%. Google, while best intentioned (I’m a fan), like damn near all companies is green washing. It’s great they’re looking for ways to reduce their energy with machine learning, but machine learning is what’s increasing their energy needs! You can’t say “we’re increasing our energy efficiency by 15% and fighting for sustainability” while working towards increasing your absolute demand by upwards of two hundred percent!
Google is working on correcting this, but no matter what, their energy demand is going up. If I ask Google Home to play a song for me rather than pick one on Spotify or my downloaded music or my girlfriend or not play music at all, that’s more energy. There is a reason Siri / Alexa / Google need an internet connection. They need access to those data centers.
I don’t need an internet connection to write down a reminder on a white board, but now I use Siri to write reminders down (because it’s less metabolically costly).
I used to walk outside more often to learn the weather. Now I might Google it first.
When I went to Europe, rather than do the best I could translating, I used Google Translate, which is based on machine learning.
And on it goes.
but while a machine will be driving the car instead of us, (most of the time) we will still be in the car. It’s not humans or machine. It’s not 96 watts or 250 watts. It’ll be humans + machine. It’s 96 watts or 250 + 96 watts. Even more likely: humans + machine + whatever other machine we’re entertaining ourselves with.
-> If self-driving cars add to the traffic problem, which people like Elon Musk think it will –Algorithms to Live By makes this argument too- then we have even more of a potential problem. I’m inclined to agree with Musk and co. See how cheap it will be to have a driver above.
It’s like when we put GPS in our cars. We didn’t add GPS energy – human energy to the car. We added GPS energy + human energy. Instead of our brains doing the navigating they went on to something else, like listening to a podcast or screaming obscenities, which may not have occurred otherwise.
Nine part series-
- AI is neat. People are messy.
- Are computers really as good as humans in chess / Go / poker?
- Classification is done by people. Not AI.
- And people are fallible.
- Liability / Are we ok with machines telling us what to do? / Loneliness
- It’s not all sunshine and rainbows.
- The more expensive the gym is the less incentive the gym has to keep us there
- Why the gym hopes you never show up.
- Improved performance methods aren’t too relevant / Improving performance filtering is a dangerous, superfluous, endeavor
- Ed Sheeran laughs at predictive analytics.
- Why you’re unlikely to find the perfect program yourself, and why AI might not be better than a trainer
- Machines can tell you what to do, but can they tell you why?
- Thinking about why other countries have more trust in their healthcare system
- The electricity bill could be insane
- Incentives still matter. Voo doo economics.
- Is any company that says it’s green, actually?
- While there are rules, like less metabolic cost, people bend them, and humans as a market are rather hard to predict
- What happened to Xbox Kinect?
Posted on April 17, 2017