According to the latest studies, power consumption per rack is increasing. What do you think about a boundary (or limit) in
per rack in near future?
This reminds me of the questions surrounding the limits of unshielded twisted pair cable a few years ago. Few would have believed it could continue to keep up with the speed of fiber, but great engineering has made it happen. I'm afraid I see per rack power consumption the same way. We're designing a small research data center right now for 25,000 Watts per cabinet. For several reasons, we've had to do it without resorting to liquid-cooled cabinets. This has only been possible only because of newer products on the market that we have been able to specially configure to do much more than the manufacturer advertises. (We've worked closely with the manufacturer, however, to make sure it will work and will be supported.) Liquid-cooled cabinets won't even do this, so, again, good engineering has made possible something that was unthinkable a few years ago. There is also no doubt that we will see liquid-cooled servers from every major manufacturer very soon. We should have them now, but everyone's afraid to be the first to make them. When that happens, 30,000, 40,000 or even 50,000 Watts in a single cabinet will not be unusual. Standard, under-floor air, however, will still be able to cool only 4 kW to 5 kW -- and then only if it is designed, installed and maintained absolutely correctly -- so it's going to take a different central cooing plant to handle these things, but it will be done. There will be even more special equipment and techniques, which will make it increasingly difficult for users to figure out what really works and in what situations, but that's been true of all technology. Let's also remember that these loads will place a new set of demands on the backup power systems. These cabinets won't wait for generators to start. Cooling is going to have to be on some kind of UPS as well, but that's a whole different topic. We'll address this in a
on high density air conditioning.