One interesting observation is the absolute power consumption levels for each platform. Historically, power consumption (and by extension thermal dissipation) has been assumed to be synonymous with TDP. While as a general guideline for a heavily loaded system, this seems to be reasonable, the margin of error is enormous. The TDP on the Dempsey processors is 130W, while Woodcrest comes in at 80W, implying a difference of 100W plus power supply inefficiency (at least 15W, likely higher). Despite this, the power draw differences ranged from 100W all the way up to 146W, quite a margin of error. The message is clear: if you actually want to know how much power you can save with an upgrade, get a watt meter, comparisons using engineering specifications are not useful.
In a modern Woodcrest based server, we observed that the power draw changes by roughly 80-100W between idle and a full load. This begs the question, how would the results change if the workload was not a ‘full load’? In reality, few if any system administrators would allow a server to be utilized above 80%. At that load level, queuing delays would dramatically increase response times beyond what most users consider acceptable. However, our performance/watt measurements (and all of those seen in the industry to date) are on fully utilized systems. Ideally, power benchmarks would scale up the intensity of the workload, showing the power consumption at different load levels. However, that is extremely difficult to do in most existing benchmarks. For now, the best way to estimate this is to interpolate using idle power draw and loaded power draw, and the utilization as a weight.
When all is said and done, it is clear that many of Intel’s claims regarding performance per watt increases are true. For the modestly configured server that we were testing, Woodcrest substantially improved the performance/watt, to the tune of 1.80-2.75x. Ironically, the improvements are most substantial for HPC-like workloads, which tend to exercise many subsystems of the microprocessor compared to other workloads. For SPECjbb2005 and XML Test 1.1, which more closely resemble commercial server workloads, the gains tended towards the low end of our observed range. The changes in idle power consumption were smaller, but still significant, on the order of 40-50W. However, it is clear that Woodcrest lives up to its claim as a more powerful, and much more efficient processor.