It’s the hardware, stupid

It is perhaps a sign of just how much attention green data centers are getting that Nature Climate Change published a report from Jonathan Koomey, Aric Masanet and Arman Shehabi  specifically focused on the characteristics of a low-carbon data center. The report itself makes an overlooked point about building a greener data center that may surprise some: that the IT hardware itself, and not infrastructure or low carbon power, is the area most deserving focus in terms of building a lower greenhouse gas emitting data center.
Much of the focus over the past decade has looked at infrastructure improvements like free cooling, hot and cold aisles, and air side economization. And increasingly the big IT companies that can afford to are looking to source their power from a renewable option, whether it’s Apple’s push into solar in North Carolina, eBay’s planned purchase of biogas offsets to make its fuel cell powered Utah data center clean, or Google’s over billion dollar push into clean power sourcing like its more recent investment in Iowa wind farms.
In terms of focusing on infrastructure, and essentially cooling, the metric of choice has become power usage effectiveness, which tells you how much of the power going into a data center is being used for the facility and how much is used for compute tasks. I’ve long felt that this metric is problematic precisely because it tells us nothing about the efficiency of the hardware itself—servers, storage and communications. The authors point out that you can have a great PUE, sending all your power to the servers, but really poor IT efficiency and thus produce higher levels of greenhouse gases. Nor does the infrastructure focus leave the door open to one day examining the efficiency of software in executing a given task, which is a far off dream right now in terms of green IT but which one day could be considered.
Still, it was interesting to see the Nature paper come out so clearly in favor of examining the IT hardware itself for its efficiency characteristics. I hope it will herald in an era of improved benchmarking for the efficiency characteristics of different server configurations and not just how to engineer a data center’s infrastructure. The authors note:

Decisions regarding when to upgrade remaining devices to more efficient models can be informed in part by a break-even analysis of the embodied emissions required to manufacture new devices versus the operational energy savings that would be realized.

This is another aspect of the calculation, one that I see people often forget about, particularly in the electric vehicle space where EVs often require more energy to produce than conventional cars. But it would be interesting for any benchmarking done that compares server efficiency to also consider the emissions involved in manufacturing new servers.
From a market perspective, this paper should bolster some of the efforts to move toward low powered chips and servers, particularly from companies like Applied Micro, Calxeda and Tilera that have long argued that both their chips and their server management software produce much improved IT hardware efficiency. We may even see a revival of the arguments from SSD proponents that there are efficiency gains to their drives versus hard disk drives. Finally the data center infrastructure management (DCIM) market which is nascent but which is precisely focused on improving utilization rates for servers will herald this type of report as proof that major gains can be made at the hardware level.
The authors make one other argument that I found interesting. They write:

Just switching an inefficient data center to low carbon electricity isn’t a good choice, because it uses up scarce low carbon electricity that could otherwise be used elsewhere.

I was struck by this statement because while clean power is definitely a very small part of the domestic energy mix, I don’t think of it as a commodity that is scarce because unlike oil or natural gas there’s no real resource limitation in say building another solar panel. The paper’s larger point is that efficiency should come first and low carbon electricity sourcing second and there are likely scenarios where if a data center uses less grid connected clean power, then more of that power stays on the grid, displacing high carbon power sourced from natural gas or coal. That’s the hope, anyways.
Thinking about the best way to build a green data center has come a long way. Now to building a system for easily benchmarking and assessing the efficiency of IT equipment!