Google Picks Gigabyte for Efficiency/Reliability
Google this week released details of how it builds servers at its major data centers. Not surprisingly, Google focused on efficiency and power, two big criteria for running an always online operation.
Google engineer Ben Jai revealed that Google used large shipping containers that can hold up to 1,160 servers at once. Each server is a 2U server chassis and are backed up by battery power in case of outages. The biggest difference however, is that Google builds each server with its own battery backup unit.
Typically, data centers have large uninterruptable power supplies (UPS) that will supply backup power to whole racks of servers. Jai however, said that this only achieves up to 95-percent power efficiency. Google's servers, which are custom designed by Jai and his team in house, use individual batteries for each server. Jai said that Google was able to achieve greater than 99.9-percent efficiency.
Using this method, Google is able to allow backup power to be supplied to servers that require it, not an entire rack. This method also allows Google to fully track power consumption and efficiency on a very granular level.
The individual servers themselves use motherboards by Gigabyte. For the past two years, Gigabyte has been positioning its motherboards as top performers when it comes down to power utilization and heat efficiency. The servers also contain two CPUs and two hard drives each, and a stock load of memory. Jai mentioned that Google uses CPUs from both AMD and Intel.
In this particular server, Google uses a Gigabyte GA-9IVDP, which is not available to the general public.
Image: courtesy of Stephen Shankland/CNET

I have been eyeballing MSI too. They slipped for a while but now seem to be offering good competitive boards with nice layouts.
I am curios about mobo model as well, but I assume it varies, just look at Gigabytes most recent server mobos. They are across the board very efficient. I suspect allot of the efficiency lies in
The 95% VS 99% efficiency quickly gets lost if you know that if 1 battery has an efficiency of 95%. 16 smaller batteries will have a total efficiency of 90,31% (every doubling of battery adds half of the efficiency to the total sum).
In other words, they'd be better off with a large capacity battery and controllers that will redirect power where needed, as opposed to separate cells for each rackserver.
Either way, google must know something Yahoo or others don't, because their search engines are just amazing!
As others have said many small batters is not more efficient. I dont konw where they are getting their 99.9% efficiency numbers, but thats completely bogus. The battery itself is less efficient then that. Unless they are only talking about transmission losses, but i really cant see them losing very much using 20 feet of wire vs 2 feet. One large supply with controllers makes way more sense, especially since you could redirect power to keep specific servers up way longer then you can if each server has it own battery.
As far as one large power supply goes. Same thing. Power supplies have an efficiency curve based on how much you draw out of it. It would be far easier to tune for a high efficiency with one large power supply then many small ones. Im sure you could get the efficiency up in the 95% range.
Its also much easier to cool the entire thing if you can isolate the power supply and backups from the servers. Pulling the power supply out of the 'server room' can reduce the cooling requirements significantly.
I mean they have hundres of thousands of servers, so im sure they know what they are doing. But in this case it looks like they chose the off the self method rather then the best method.