Sunday, March 18, 2007

MIPS per Watt: The Progression...

Last week I purchased a Kill-a-Watt Electricity Usage Monitor, which measures the wattage used by any plugged-in device. It's already proving its value.

I began measuring the electrical consumption of the three workstations I use the most. The numbers strongly suggest that the power savings from multicore are real. The question that remains is whether they're substantial enough to matter to many folks, especially small sites with only a few systems. Here we go. (Notes: All workstations are Dell systems with 2GB or 3GB RAM and 2 SATA HDDs. The performance measurements are the processor arithmetic tests from the highly regarded Sandra benchmark suite. Energy consumption measured when the systems were at rest. Systems are shown in chronological order of manufacture, oldest ones first.)


The upshot is that the dual-core systems are definitely the performance equivalents of earlier dual-processor Xeon beasts (look at the performance parity between the two Intel processors), but enegy consumption of the multicore systems is almost 40% less.

However, the difference in energy consumption between the the Pentium D and the AMD system is not that great. Moreover, the difference in CPU performance, while appearing to be a lot, feels the same when I'm at the controls.

So, I think multiprocessor boxes are easy candidates for replacement by multicore systems, but upgrading multicores does not look compelling currently. (The Pentium D system is about a year older than the AMD system.)

2 comments:

TimH said...

Hold on - surely this is only valid if the energy consumption is measured when it is actually working at that CPU level? For all I know, the newer dual-core machines, for example, might triple their power consumption when going flat out, turning your results upside-down...

Andrew Binstock said...

Tim: True enough, but the average workstation spends much of its time doing very little. Going flat out would provide possibly different numbers, but they would reflect the usage for only a very small time slice of the machine's workday.

A ratio I've seen used by power estimation software is seven hours per work day at rest (but not hibernating), and one hour of serious processor work. I think that for most business desktops, this is still too generous (that is, they work even less).