AMD and Intel differ on Energy Star server specifications

Just how green should servers be?

In response to prompting from the US Congress, the Environmental Protection Agency (EPA) is attempting to draft a new Energy Star specification for enterprise servers, but is having some problems agreeing on terms with the IT industry.

AMD, Intel, IBM, Dell and Google have all responded to the Environmental Protection Agency's (EPA) already revised, first draft of the Energy Star specification for enterprise servers. Not surprisingly each company has its own agenda and there is little agreement in their responses. This harks back to the original Energy Star initiative, when critics said that industry lobbying was so successful that almost no changes were required to existing designs to meet the Energy Star criteria, as eventually published. Lobbyists claimed that the changes required by the original Energy Star proposals would have been so large that they would have crippled the industries involved.

Following an April meeting with industry representatives, the EPA invited written responses to the first draft of the specification from the participants and these have now been submitted and published. In places they differ so widely in their conclusions that reaching an agreement on the new Energy Star specification is not going to be easy. Broadly speaking none of the companies involved can even agree on what constitutes a server, or on how the power efficiency of servers should be tested. The EPA has responded to some of these arguments by publishing a revision of the proposed standards definition of "enterprise server" – PDF file. To have any impact on gross power consumption the eventual standard must relate to the majority of machines actually in use as servers.

Advertisement

Part of the problem is that there are systems, like 'blade' servers, that are obviously designed as servers and then there are desktop systems sold, or used as servers, which may, or may not, have design features, for example extra processors or large RAIDs, that make them suitable as servers. Perhaps the simplest definition would be to say that a server is any system running a server operating system and functioning as a server, but for the EPA this isn't a particularly useful definition because, for the purposes of a power consumption standard, the EPA are assuming it is the characteristics of the hardware that matter. The relative power efficiency of different server operating systems in their interactions with hardware, is apparently, being ignored.

Google, an end user of servers rather than a hardware manufacturer, takes a broad view on the definition of a server. In their response, Bill Weihl says "The definition proposed for servers seems unduly restrictive." and "...to capture the 'volume server' part of the market, it seems important to eliminate as many restrictions in the definition as possible."

AMD and Intel, of course, focus on how the new Energy Star specification will favour the strengths and weaknesses of their CPUs and chip sets. The two chip makers do at least seem to agree that a distinction should be drawn between systems based on the number of processors supported and say that systems with more than four processors should be the subject of a separate specification. Although there is some imprecision in the documents over whether the number of CPU sockets, or the number of CPUs actually fitted, are being discussed. AMD's reply mentions only processors while Intel's refers to sockets.

However they disagree over whether or not the SPECpower_ssj2008 power efficiency benchmark – which AMD helped to develop – should be a component of the Energy Star specification. AMD say it should not, as this would "conflict with the spirit of the release guidelines of the SPEC" – Dell seems to agree with this view. Intel is in favour of including SPECpower_ssj2008, probably because the benchmark is only based on a single workload, e.g. a server-side Java application and the Intel Xeon processors perform well in this situation, currently topping the SPECpower_ssj2008 results list. AMD is already working on optimising its Opteron processors for Java.

Intel does take issue with the idea of taking into account the power consumption of an idle server in the efficiency assessment. There is a concrete technical and perhaps biased reason for that opinion: current Xeon systems generally use fully buffered (FB)DIMMs, which consume comparatively high amounts of power, even when a system is idling.

IBM prefers to define the number of processor sockets as "one or more", and want to include plug-in processor boards.

The background to all this is that in late December 2006, the US Congress called for the Environmental Protection Agency (EPA) to take measures to control the rapidly growing energy demand of computing centres. The power that these consumed in the United States in 2006 made up around 1.5 per cent of total electrical energy consumption. The EPA has been working with the IT industry since August of 2007 on an Energy Star specification for enterprise servers and presented its first draft of the specification – PDF file – at the end of February 2008.

Historically, commercial imperatives have proven to be more powerful than legislative controls. Arguably, the pressure of the now steeply rising cost of power is likely to make truly power efficient server products much more commercially attractive to the customer, regardless of Energy Star guidelines. The problem with Energy Star has always been that it is not a comparative rating, but simply sets maximum levels that compliant equipment should not exceed.

(Terry Relph-Knight)

(trk)