Google Testing AI System To Cool Data Center Energy Bills 52
An anonymous reader writes: Google is looking at artificial intelligence technology to help it identify opportunities for data center energy savings. The company is approaching the end of an initial 2-year trial of the machine learning tool, and hopes to see it applied across the entire data center portfolio by the end of 2016. The new AI software, which is being developed at Google's DeepMind, has already helped to cut energy use for cooling by 40%, and to improve overall data center efficiency by 15%. DeepMind said that the program has been an enormous help in analyzing data center efficiency, from looking at energy used for cooling and air temperature to pressure and humidity. The team now hopes to expand the system to understand other infrastructure challenges, in the data center and beyond, including improving power plant conversion, reducing semiconductor manufacturing energy, water usage, and helping manufacturers increase throughput.
Re: (Score:2)
Your confusing General AI with weak AI.
General AI we are a long way from achieving, however specialized domain specific AI is something we've been doing well for a while. We're talking things that can analyse data and make decisions not programmed into it but by using learning algorithms.
Re: (Score:3)
So by weak AI we've had AI since the first programmable device?
Re:God damnit (Score:5, Insightful)
So by weak AI we've had AI since the first programmable device?
No. There's a very large difference between machine learning-based systems and human-programmed systems. One very obvious one is that human-programmed systems can be explained by pointing out the specific rules that are being applied, in what order and for what reasons. With machine learning, especially the deep neural networks that are currently proving so effective, we can explain the structure of the system and the mechanisms used to "train" it, and how inputs flow through it to produce outputs, but we can't explain what logical "rules" are being applied, or when or why.
In theory, it should be possible to derive the rules by examining the weights in the neural network. In theory, it should be possible to implement those rules with traditional, human-defined logical rules. In practice, the details of how the trained system works are usually beyond us, and typically better than we can do with explicit logic.
Re: (Score:2)
Kind of sounds like the difference between traditional and cloud based servers. With traditional server you know exactly what you are dealing with, with cloud based there could be any combination of resources being used to address your needs (hopefully in the most efficient way possible). Probably the main addition being self adapting.
Re: (Score:2)
Also to add, IMHO, I feel like Mass Effect had a better definition with AI vs VI. AI being capable of reprogramming itself to do new and unexpected things, where as VI was constrained by its software for specific purposes and even certain limits within those purposes.
Re: (Score:3, Insightful)
Re: (Score:3)
Here is a hint: a neural net isn't anything like a brain.
There are some vague structural similarities between a neural net and the basic structure of the brain, but they're certainly very different, yes.
And now they just add "Deep" to the front of "neural net" or "learning" and think that is must be, like, really good now, because it is "deep" learning.
You don't have any idea what the "deep" means there, do you? It's not a content-free buzzword, it has a very specific meaning... and you might consider what the fact that you don't know what it is says about your knowledge in this space.
We aren't any closer to AI than we were 40 years ago.
Right. We've made, for example, no progress at all on voice recognition.
Re: (Score:2)
Plenty of AI is not based on 'machine learning', e.g. playing chess.
Bottom line AI -- or what is AI -- is defined by researchers working in that topic, not by general masses.
And so far away from 'general AI' as many people think, we aren't anyway, look e.g. at IBMs 'Watson'.
Re: (Score:2)
Still it isn't really AI, it is an adaptive system.
Where it just calculates correlations from existing data and chooses what would be the more optimal solution.
It is just more advanced than a simple workflow.
Re: (Score:2)
That funny because I am so sick of people complaining of the allegedly incorrect use of the word Artificial Intelligence.
It is the correct academic term for the computing field that include data mining, machine learning, operation research, and statistics. Now, maybe Artificial Intelligence does not mean what you wish it meant. But academics are fairly clear on what it means...
Re: (Score:2)
Re: (Score:2)
Artificial Intelligence is a field of study, just like Algebra.
The way ACM defines it [1], AI includes computer vision, NLP, planning, and knowledge representation and reasoning. All these tasks are today based on computational statistic/machine learning. You can also look at the last AAAI conference [2].
The way it is defined, your GPS is an AI system.
[1] http://dl.acm.org/ccs_flat.cfm [acm.org]
[2] http://www.aaai.org/Conference... [aaai.org]
Re: (Score:2)
The GPS certainly not.
The map plotting/route finding: yes.
Re: (Score:2)
So what do you call artificial creations that demonstrate (limited) intelligence? No one is claiming this thing passes the Turing test.
Re: (Score:1)
Is this deep mind application the world's fanciest (Score:2)
"Machine that turns itself off!" ? (thus saving a lot of AI server power)
https://www.youtube.com/watch?... [youtube.com]
Re: (Score:1)
I wouldn't be able to tell if it were a human configuring the data center or an AI simply because of the limited chances to express agency.
Wow (Score:2)
Google Thermometer! (Score:2)
Google has found another earth-shattering ML application: Google Thermometer! For a fee of only $xxxx (fill in the blanks), Google will monitor your a/c system and control it remotely! It will also accept Twitter, e-mail and messages from irritated occupants (such as Mary in Accounting, who is enduring post-menopausal heat flashes) and act (or not) upon them.
Next Month's Fun: Google Rectal Thermometer, which will monitor your bodily functions to optimize your metabolism.
The real link (Score:2)
DEEPMIND AI REDUCES GOOGLE DATA CENTRE COOLING BILL BY 40% [deepmind.com] (I can't see a fixed ink, and its DeepMind who is doing the shouting)
It really pisses me off when I have to jump through hoops to find the actual guts of the story rather than someone else's opinion of the story.
Re: (Score:2)
Even from that link it's not clear how they actually achieve energy savings.
There are a lot of software buzzwords, but no real description of how they can reduce
energy consumption. ("Ye cannae change the laws of physics...")
And the single graph has no units on the axes.
Re: (Score:2)
Even from that link it's not clear how they actually achieve energy savings
There are a lot of software buzzwords, but no real description of how they can reduce
energy consumption. ("Ye cannae change the laws of physics...")
I agree that its light on details but at least I can get a sense of what they are doing is optimizing how and when their cooling systems run based on mining historical data for patterns that are non-obvious to humans and provide more efficient controls than simple control system feedback loops.
And the single graph has no units on the axes.
I don't mind the lack of units. A 40% differential between before and after shows that they are making a hell of an improvement over current systems.
This Adds to my Nest Frustration (Score:1)
Re: (Score:2)
Re: (Score:2)
Wouldn't virtualization make this pretty trivial, concentrating your workloads on hosts served by a single zone?
As visually appealing as it is, I've kind of always wondered why so many data centers I've seen have been basically one giant room. Maybe they have some zoning control with the whole hot aisle/cool aisle concept, but it kind of seems like if they were more segmented there would be easier zoning controls and less general air mass cooling to be done.
Although I would imagine data center cooling is a
Re: (Score:2)
Depends on what you mean. It makes reorganizing the machines pretty trivial, potentially, assuming they all have similar hardware capabilities (which may or may not be a safe assumption). But there is a cost (both in power and time) to moving large quantities of data from one machine to another, so you can't necessarily shift load around instantly, depending on what the machines are doing and ho
Re: (Score:2)
I'd wager that's the value of the AI analysis, sifting the performance demands across the variables -- I/O, CPU, etc -- to find patterns that indicate where you ought to home workloads, the return on migrating workloads on demand and so on.
It may be that the metrics they had targeted previously suggested similar kinds of workload distribution merely for the power benefits of idling or spinning down nodes but without taking into account the HVAC-related aspects, resulting in fewer nodes but with no HVAC zoni
The obvious (Score:2)
There's surely a Nest joke in here somewhere...
Cool bills? (Score:2)
If they want to cool their bills, I recommend they invest in a simple fridge and stick them in there. I find it keeps my bills quite cool - if I bother to keep them in the fridge at all, of course. For some reason it only rarely happens...
Here's the Source (Score:2)
Ugh (Score:2)
Scripts that watch temperature parameters doth not make an Artificial Intelligence.
It's a sensationalized way of saying Google finally put a programmable thermostat into their data centers.
( " Joe wrote a script to control the AC " doesn't quite come across with the same flair now does it ? )
When someone finally DOES give berth to a legitimate AI, we're going to have to redefine the term I think because
comparing Google's new thermostat to a true AI system would be akin to comparing a match head to the GD Su
Adaptive HVAC control (Score:2)
Been around for a few decades.
Google has the advantage over office buildings, schools, etc. in that the amount of energy consumed by data centers is so large that the cost of additional sensors can be justified. And the more inputs you have, the 'smarter' your controllers can be.
Re: (Score:2)