Power Consumption And Noise Comparison
We used Ubuntu Server 10.10 and Folidng@Home for power consumption and acoustic testing. It's fairly common to see these measurements taken fairly quickly after a configuration is presented with load. But instead of taking a reading 15 minutes into a workload, we let these servers run at 100% utilization for six hours each. This actually yields an appreciable difference in power consumption and noise compared to the shorter testing used on desktop hardware, and is probably closer to what a server might see in a data center. We picked Folding@Home because it is a real-world GROMACS-based application used for research purposes, and a big work unit takes 24 hours or more for the machine to complete. It also scales fairly well from four to 64 cores, making it useful for generating high CPU load over long periods. Both Linux and Windows clients are available, but the Linux client tends to be a quite a bit faster on the same hardware, and Linux gives options for better thread management.
Power consumption and noise testing was performed using each motherboard's PWM-controlled fan settings. This is worth noting because there is one school of thought that suggests you should keep fans at full speed in order to push as much air through the chassis as possible. A side benefit of pegging fans at full speed is that there is no ramp up or down, putting additional wear on the fans. The other school of thought, and one that I ascribe to, is that motherboard PWM fan control is good enough to keep components cool while saving power. If workloads are relatively even, there is not much ramp up or down anyway, allowing the coolers to spin at lower average speed. Just to give you an idea, at full speed, the fans used in these barebones units can draw between 15 and 18 W each. With four to eight fans per system, that's not a negligible amount of power. Data centers typically have restrictions on power delivery, so cutting back may help save overage charges.
An Extech 380803 True RMS power meter was used to measure power consumption of each system. Measurements were taken at the six-hour mark and averaged over three runs. The systems were allowed to cool for three hours between runs in a fully powered-off state.
Supermicro and Intel achieve almost identical power consumption. Tyan isn't far behind. However, there is a measurable delta between the three systems. There is a distinct possibility that the difference is due to Supermicro and Intel using 80 PLUS Platinum-rated PSUs, while Tyan arms its machine with less-efficient Gold-rated power supplies. Also, Tyan uses a separate PCB to distribute power around the chassis.
For sound testing, we used an Extech 407764 NIST-calibrated digital sound meter with a 15-foot microphone extension cable so that we could monitor from another room to preserve a 32 dB(A) sound floor (the Extech 407764 is rated for 30-130 dB sensitivity.) The idle measurements were taken 30 minutes after the system booted into Ubuntu Server, and the load readings were taken before each Folding@Home six-hour run was completed. Three runs were averaged for the scores below.
Intel does well here, as does Supermicro. The Tyan chassis is a bit louder (particularly at idle). But being a 2U enclosure, there is significantly less noise than if these were 1U machines. All three servers are too noisy to put in the middle of an open office or in a home. However, in a data center, these noise levels are expected.
Realistically, these machines will run under at least some load most of the time. So, while idle noise is recorded for the sake of completeness, the results are less relevant in the real world than the load numbers.