I’ve been curious for a while about how much power Opteron CPUs draw when idle, so I set aside a bit of time to do some measurements. I used a Supermicro 1U system with redundant power supply. The motherboard model is H8DGU-F. The system has 32GB of DDR3 ECC ram, and two Intel X25-M 120GB SSDs. There are two Opteron 6128 CPUs installed. These Opterons have 8 cores each, and they run at 2.0GHz. These are the CPU power specs:
Average CPU Power | 80W |
Thermal Design Power (TDP) | 115W |
The ‘Average CPU Power’ is based on ‘average’ use, which is explained on Wikipedia.
According to Wikipedia, the Thermal Design Power is the maximum power consumption for thermally significant periods running worst-case non-synthetic workloads (cf. this article). If we assume that the bulk of the electrical power consumed by a CPU is converted into waste heat, then the TDP can be a reasonable approximation for the amount of electrical power a CPU would consume under a worst-case, real-world load.
I used cpuburn to generate such a load. There was no IO load on the system during the tests. I measured power draw with an off-the-shelf Kill-a-watt, so these results should be taken with a grain of salt.
16 cores | idle | 145W (153VA) |
8 cores | loaded on 1 cpu | 215W (221VA) |
8 cores | loaded spread over 2 cpus | 235W (243VA) |
14 cores | loaded | 277W (285VA) |
16 cores | loaded | 290W (297VA) |
The data indicates that the idle vs. full load power consumption difference for one CPU is 70 to 75W.
If we assume the power consumption under full load is 115W (the TDP for the processor), then idle power consumption would be 40 to 45W per CPU. That would put idle power consumption at 35-39% of its TDP for this particular CPU.
You might want to test with a build prior to the power regressions introduced in 6.38 (not sure if it was bisected yet) and see what the difference (if any) is. From what I’ve seen, it might be significant.
Chris – I ran these tests on (the Xen kernel of) Debian Squeeze, which is a 2.6.32 based kernel, which does not have that regression AFAIK.
Indeed, 32 doesn’t have that regression (it’s pretty recent).