What's cool about DC-powered data centers

My recent story on DC-powered data centers seems to be getting some mail (see Chilling Out With DC Power). That feature discussed how using DC power might reduce heat loads from energy inefficient AC to DC power conversions in densely populated server racks and increase the reliability of the power delivery system in data centers.

Jim Liles, who works for an engineering firm, e-mailed to ask whether there is enough real efficiency difference between DC and AC power supply systems to have a significant impact on air conditioning requirements. That’s the $64,000 question, as the experts I spoke with disagreed on this point.

“If their was enough real efficiency difference between the DC-to-DC and the AC-to-DC to have a significant impact on air conditioning and densities we would take a real hard look at this approach,” he says.

The answer, I suspect, is that it depends.

Jim points out that if you deliver –48v bulk DC power to the data center you still have to step it down within the rack to the individual voltages required for each server and its components. With regard to the rack-level DC power supplies, Rackable Systems claims that those have the exact same conversion efficiency as does the bulk DC power delivery system - about 93%. That means your end-to-end total efficiency loss would be somewhere around 14% versus about 30% for a standard AC power supply attached to a 1u server.

Jim also points out that –48v DC power requires a conductor that’s 17 times larger than what’s needed for delivering the 480 volt 3-phase AC power into the room and 7.5 times larger than what’s needed for 208 volt 3-phase power typically used in floor-mounted distribution centers. The argument made in the story was that those costs tend to be offset by savings in other areas.

Jim mentions that in the story IBM says there’s only a 3% net efficiency difference while others claim as high as 50%. The reason is that the efficiency depends on the load and power supply design. But IBM is claiming those higher efficiencies only for its BladeCenter power supplies because they power up to 14 servers within a single chassis. That means total utlilization is higher, which increases efficiency. But for most power supplies on 1u or bigger servers there’s a lot of idle time that greatly reduces efficiency – especially when redundant power supplies are used. For regular 1u servers, IBM cited 70% efficiency as more the norm.

I asked the engineers at power supply vendor Astec Power about AC power supply efficiency. Depending on the scenario their estimates ranged from 84% to 92%. When asked what are the likely savings from using the DC power approach a spokesperson said: "You save the AC to DC boost converter (400vdc), but add much higher switching currents to the secondary stage, so it is a wash. Also, the mix is currently 20% DC input, 80% AC input even when there is also a telco -48v requirement."

Others had different opinions. For example, Geoffrey Noer, Sr. Director of Product Marketing at Rackable Systems, says that while DC power distribution units are more expensive than AC systems, “cabinet level costs” for AC are slightly higher because of the need for all of those AC power supplies. “I’m not convinced it’s that much more difficult to do DC designs,” he says. But that begs the question: if it’s so easy, why aren’t more organizations doing it?