Slashdot Mirror
Immersion Cooling Drives Server Power Densities To Insane New Heights (datacenterfrontier.com)
1sockchuck writes: By immersing IT equipment in liquid coolant, a new data center is reaching extreme power densities of 250 kW per enclosure. At 40 megawatts, the data center is also taking immersion cooling to an entirely new scale, building on a much smaller proof-of-concept from a Hong Kong skyscraper. The facility is being built by Bitcoin specialist BitFury and reflects how the harsh economics of industrial mining have prompted cryptocurrency firms to focus on data center design to cut costs and boost power. But this type of radical energy efficiency may soon be key to America's effort to build an exascale computer and the increasingly extreme data-crunching requirements for cloud and analytics.
"INSANE" new heights?
Has Slashdot been sold to the Gawker network now?
-Styopa
...they want their cooling back: Cray-2
I was actually one of the admins for a Cray-2 (and other systems) at NASA LaRC from 1988-1992. It was pretty cool (no pun intended). The chassis was Plexiglas (or something else clear) and you could see the 3D circuit boards immersed in the Fluorinert - which was wicked expensive back then. I always wanted to put some plastic fish inside the system... The system was moved to the Virginia Air and Space Center (VASC) for a while after being decommissioned sometime later.
It must have been something you assimilated. . . .
The difference with this approach is two-phase cooling, where they're actually boiling the heat transfer fluid. That can remove heat a lot more quickly, as long as you can keep a few issues under control:
1) Getting a working fluid with an appropriate boiling point and otherwise acceptable physical parameters (non-flammable, doesn't dissolve your circuitry, etc). 3M has already stepped up to the plate on that.
2) Recondensing the vapor fast enough. This is a lot easier than cooling the circuits directly.
3) Preventing the hot chips from forming a vapor barrier, which insulates the chips from the coolant. The Leidenfrost effect is an example of this, but you can lose efficiency long before you reach the droplets-skittering-around level, especially if there are lots of nooks and crannies where bubbles can get stuck. Presumably the designers have handled this as well.
If they go with a transparent enclosure and some gratuitous lighting, this could become the new mad-scientist/Big Scary Computer visual trope. Let's face it, lab coats, blinking lights and reel-to-reel tape drives are really tired...
Better hurricane models (the current prediction technology had the worst hurricane on record cause -zero- deaths in Mexico.)
That was caused by sensational journalism rather than bad models. Although Patricia was a record storm out at sea, models showed that it would lose energy as it passed over cooler waters close to the coast. The models correctly predicted that the winds would drop from 200mph to about 165 by the time it came ashore.
The next big use for large machines is HFT. A microsecond or two on a fast pipe can mean millions in stock gains.
The glory days of HFT are in the past. Speed is no longer an advantage when everyone is doing it. Besides, HFT needs fast pipes, but doesn't really need a lot of computation.