Cooling with oil? Not exaclty new
Green Revolution Cooling have been at this stuff a while now. (I do not work for, or am I affiliated with GRC)
Wouldn't fancy "servicing" a server though, messy me thinks.
Facebook is dunking its servers in gloop in a salt shed in Oregon so it can overclock their processors, The Register has learned. The experimental, brutal "immersion cooling" scheme was revealed to us by Facebook hardware design and supply chain bigwig Frank Frankovsky on Friday. Though Facebook had tested out an immersion …
How much CPU time does it take to generate a typical facebook response ? How significant is that CPU time compared to disk/network delays ? What is the cost (build & maintainance) of putting machines in oil ? Would they not be better just installing a few more machines ?
Overclocking is good for long running heavy CPU situations where extra paralellisation does not help; weather forcasting comes to mind.
When you have billions of facebook responses to generate, your servers are going to be running at high capacity. That's why FB needs entire datacentres in the first place... if they have 1000 servers and overclocking can gain 20%, that's a lot less servers.
Plus, this is just R&D.
The CPU time won't be used to generate simple HTTP responses. It'll be used for face recognition and other image processing so that they can target advertising better.
Half the photos of you also include your car - sell you motoring stuff. Most of your photos are of you outdoors - sell you hiking boots and waterproofs.
That may be an unexpected benefit of the "fake capacitor" fiasco of a few years back - remember all those bulging/bursting electrolytics which used a (badly) copied formula? I've seen a few motherboards lately (like Gigabyte's "Ultra Durable" range) bragging about using solid polymer capacitors, rather than traditional electrolytic ones with liquid electrolyte vulnerable to the problem you mention.
Besides, when you're buying on Facebook's scale ("We're thinking of filling our new datacentre with your motherboards, will you help us get them working in liquid coolant?") the suppliers have a rather bigger incentive to help than if you or I asked the same question about a home system: I'm sure they'll know, or investigate, what it takes to do this properly. Running a few test systems for years wouldn't bother them, and the component manufacturers probably have a fair idea already: immersion and liquid cooling's not just for PCs, after all.
Viscous gloop isn't going to be that great as a coolant. The viscosity just slows down the convection needed to carry the heat away. You need a thin liquid, or even better, a thin low boiling point liquid with a large latent heat of vaporisation. In that case the liquid boils on the hot components. That technology has been used for decades for high-power thermionic valves.
Such a phase change, (liquid to gas), can carry large amounts of heat away from hot components, but suffer from at least two drawbacks.
1. Once the coolant is in its gaseous phase it becomes a rather efficient insulator, preventing liquid coolant from having maximally efficient contact to the hot component.
2. The collapse of the bubbles of gaseous coolant can cause the erosion of the surfaces adjacent.
A viscous coolant which phase changes to a more fluid liquid solves both of those problems.
I don't know if that's what they're doing but it's a possibility.
Normally, overclocking involves boosting voltage, not just cooling. This leads to increased electromigration, which won't be affected by cooling, and is a major cause of failure under overclocking conditions. I wonder if they're manipulating that aspect at all.
For a better approach, why not use Ethanol and Dry Ice? (And, if things get too bad, the techies can resort to drinking the coolant!). Note that this isn't unprecedented (Err, the Ethanol and Dry Ice, not necessarily the techies drinking it, although that has been known to happen, too!).
There are some mainframe manufacturers who have resorted to using Helium as a cooling agent (e.g., "TCM"). Plus, there has been some work done on using Liquid Nitrogen immersion as a cooling agent for overclocked systems.
Large electrical generators are frequently cooled by Hydrogen gas (It has a low viscosity, so as to not interfere too much with the rotating components, and conducts heat well. Of course, one has to have a very good seal on the bearings, else one runs the risk of the "Hindenberg Effect".).
P.S. I'll get my coat; it's the one with the Kentucky bourbon and Dry Ice in the pocket.
Cold improves transconductance, which used to make logic gates faster.
But cold also increases the threshold voltage, which has a bigger effect these days. So you want your overclocked machine running hot for max speed. This is going to make electromigration even worse though.
Biting the hand that feeds IT © 1998–2021