>Our machine learning system was able to consistently achieve a 40 percent reduction in the amount of energy used for cooling, which equates to a 15 percent reduction in overall PUE after accounting for electrical losses and other non-cooling inefficiencies
So the actual savings were 15%? Which is still significant for sure. I'm guessing their next step will be to reduce those non-cooling inefficiencies.
This is pretty stunning and should make a lot of knowledge workers terrified. I'm sure Google has a bunch of really smart guys working on figuring out how to reduce energy costs. To have a computer come in and get those kinds of results on the first attempt is pretty mindblowing.
I would counter the fear. Data center engineering requires expert knowledge (network, civil engineering, material science, etc), and expert knowledge is a testimony of experience, and experience is only proven to be good with data (track of records!)! While computers are very good at making prediction and decisions with enough data, humans are still integral in the decision such as actually building the blueprint and architecting the right system. That computer cannot do. Only human is capable of doing that today. Not to be too negative because I don't know how complicated the model is, but since they already have senors all the years (they use these senors to determine how much cool air need to be come in, how much of hot air need to be released and reused, etc), I am not sure how the learning model is different from a classical supervised ML algorithm. I can only think of simulation wasn't in the original equation and here they do some simulation?
"I am not sure how the learning model is different from a classical supervised ML algorithm"
It's different because you have to perform a sequence of actions that optimize some goal. This is not a classical prediction/classification problem, it is a reinforcement learning problem.
This is why I prefer the nonspecific "meat sack," though I'll admit mine is often more cumbersome in general usage. (Seriously: "folks" is what I tend to use to avoid this situation. Any time you find yourself saying "guys," folks will work.)
If you s/guys/folks/ in that comment and think it sounds like a politician, I'm quite intrigued by who your politicians are. Kill this thread and don't reply to me, though, we are super off topic.
While AI is not at a point to write programs, it's probably going to eliminate the need for many programs.
A majority of programs exist in order for people to use them to achieve some result. If the AI can achieve the same result, you've eliminated the need for the program, and the programmers who would program it, and the people who would use it (and the sales people who would sell it).
Well, to be fair, the AI probably tried somewhere over several millions of times before coming to whatever conclusion it reached in its training process
When people talk about AI, the topic invariably turns to things like "truck drivers and other low-skill jobs should be scared!" I think the reality is the opposite. I think that anyone whose job is to do mid-to-high skill work with numbers should be scared, because you can automate their jobs without solving the messy problem of getting computers to move big physical objects around.
This wasn't clear to me either. The best I could understand was from this statement
"The purpose of these predictions is to simulate the recommended actions from the PUE model, to ensure that we do not go beyond any operating constraints."
Seems that the action was to deliver the perfect amount of cooling needed for the current/predicted load. I'm guessing the old model was sloppy with the cooling and used more power to keep the servers cooler than they needed to be.
Yep, all systems until literally ideal will have efficiencies to iron out. Unless your data center is identical to Google's, taking any specifics to play would likely be ineffective advice!
Always fascinating how it turns out many ideas in research were already tested in the past (of course DeepMind is most likely aware of and has improved on prior work).
This makes me wonder what a next-gen datacenter would look like; far more control points, far more data gathering, probably some combination of high- and low-inertia cooling with different characteristics -- essentially you are going to want to start giving your NN more knobs to tweak.
That seems very singularity/jackpot-ish to me. Cool, but will have some interesting unexpected consequences, I bet.
AI is a field of research with the goal of creating a computational model of human cognition, that's the scientific part. The engineering part is then to build a system based on the computational model such that this single system can be used as a general purpose problem solver.
yeah, if it works; it is no longer AI. It is like a rainbow or a horizon—it should be always just out of reach otherwise it is not a real thing (according to some people).
"We are planning to roll out this system more broadly and will share how we did it in an upcoming publication, so that other data centre and industrial system operators -- and ultimately the environment -- can benefit from this major step forward."
Please do, I am hungry for details here. The one chart they put in has no scale on it's axis, and as another comment pointed out, they didn't give any details about what recommendations were followed to achieve the improvements.
40 comments
[ 2.5 ms ] story [ 83.9 ms ] threadSo the actual savings were 15%? Which is still significant for sure. I'm guessing their next step will be to reduce those non-cooling inefficiencies.
It's different because you have to perform a sequence of actions that optimize some goal. This is not a classical prediction/classification problem, it is a reinforcement learning problem.
While AI is not at a point to write programs, it's probably going to eliminate the need for many programs.
A majority of programs exist in order for people to use them to achieve some result. If the AI can achieve the same result, you've eliminated the need for the program, and the programmers who would program it, and the people who would use it (and the sales people who would sell it).
"The purpose of these predictions is to simulate the recommended actions from the PUE model, to ensure that we do not go beyond any operating constraints."
Seems that the action was to deliver the perfect amount of cooling needed for the current/predicted load. I'm guessing the old model was sloppy with the cooling and used more power to keep the servers cooler than they needed to be.
From NIPS in 2008: "Managing Power Consumption and Performance of Computing Systems Using Reinforcement Learning" (http://papers.nips.cc/paper/3251-managing-power-consumption-...)
That seems very singularity/jackpot-ish to me. Cool, but will have some interesting unexpected consequences, I bet.
Very soon after that, the AI will be revisiting the stored history of the internet, will be reading this post, and laugh its ass bits off...
HELLO FROM THE PAST! WE WERE HUMANS!
Please do, I am hungry for details here. The one chart they put in has no scale on it's axis, and as another comment pointed out, they didn't give any details about what recommendations were followed to achieve the improvements.