DeepMind AI bots tell Google to literally chill out: Software takes control of server cooling DeepMind’s artificially intelligent algorithms are directly controlling the cooling systems within Google’s data centers to improve efficiency. Google’s vast warehouses are filled with hundreds of thousands of heat-generating servers, chugging away to keep all those adverts, Gmail messages, YouTube videos, and Google Search …
...we need AI to implement a thermostat nowadays?
(shuffles off, muttering about the youth of today, and how discrete electronics should be enough for anyone, all started going downhill with those fancy-pants op-amps, don't get me started about when microcontrollers came along, and now everything needs plugging into the internet to work, not that anything turns on instantly any more, oh, no, in my day... cont. p.94)
Themostats suffer from delays which makes them inefficient. If heating, or in Google's case cooling, is sufficiently expensive the investing in efficiency makes sense. A simple, rule-based approach might be sufficient. For example, last year I swapped the simple radio themostats in my flat for ones with clocks and schedules and was able to keep the flat at a more agreeable temperature as a result with less manual intervention, even though the principle of the thermostat remained the same. I don't have sufficient data but I think the rule to throttle back overnight and set holiday timings has reduced my heating bill while increasing comfort (especially when returning from a long trip during a cold spell).
The rules could be extended using measurements such as termperature and humidity, amount of daylight, etc. This makes it a good domain for machine learning but still requires people to select features and outcomes, something that itself can be optimised in certain environments. Clearly not relevant for my environment but for a company the size of Google the savings could be relevant and it's also an excellent test case and advert for the technology. Because, despite the savings, this is largely PR.
You don't need to go into the realms of Star Trek to find magnetic bearings: they are already used in some high performance systems. As for mechanical thermostats, even back in the 80's it was a pretty crude industrial application that didn't go a step further and use electronic PID control.
And are allowing unloaded parts of the datacenter to run without cooling for a while, or similar.
I can see value in being able to automatically switch between cooling systems, and to reroute cooling away from unloaded segments.
Otherwise this would be nothing that a decently tuned PID controller wouldn't do cheaper and better.
The value is probably in predictive analysis - ie trend data to optimise pre-cooling and moving to virtual predictive 'simmerstat' rather than thermostat approach combined with any optimal combination of cooling techniques (free cooling, thermal mass, solar, ice making etc).
'decently tuned PID' - getting the decent tuning is that hard bit - and long an AI application in its own right. You also have to managed a network of the things that add noise and complex interactions. That takes significant investment of domain engineers and is rarely worth it in complex systems (like commercial buildings). You also have to manage context drift on those tuned PID controllers.
There are benefits to be had from explicitly doing the predictive analysis (in some senses a PID controller is a simplistic implicit predictor) and a separate constraint based optimiser. That can be done at the system level and with set point adjustment thus be reasonably robust with badly tuned local controllers.
The interesting bit is whether we have reached the point that this can be done with general tech and general it skills and thus save energy rather than every time being a engineering problem. Its long been possible to improve heating system efficiency and add simple predictive controllers - just rarely has it been a sustainable use of the necessary skills.
In data centers of course getting smart whilst being able to rule out over set point overshoot is probably worthy of study all in its own as staying with the thermal design parameters of the kit is rather important ...
The mechanical thermostat has hysteresis, slow(ish) response and mechanical failure modes.
So more likely a network of thermistors feeding a control station even without the AI.
TBH though I would have thought the problem is sufficiently constrained in dimensions that something rather simpler would have done.
OTOH if you do have all that "Deep learning" stuff set up maybe re-purposing it is pretty cheap (and as we know to the man with the hammer everything is a nail).
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
