TDP as power used?

[CRGreathouse]

I'm interested in comparing CPUs in terms of (among other things!) power consumption. How reasonable is it to assume that power consumption at full load is equal to TDP?

[bgbeuning]

From my experience, prime95 is usually limited by something other than
the CPU so the CPU is not 100% busy and does not use all the power it can.

Here is my one example for an i5-3470 rated at 77 TDP

With 1 DIMM, the entire system uses 67 W and gets these benchmark results

Timings for 4096K FFT length (1 cpu, 1 worker): 24.78 ms. Throughput: 40.36 iter/sec.
Timings for 4096K FFT length (2 cpus, 2 workers): 31.09, 31.08 ms. Throughput: 64.34 iter/sec.
Timings for 4096K FFT length (3 cpus, 3 workers): 41.77, 41.73, 41.81 ms. Throughput: 71.82 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 56.00, 57.24, 56.07, 56.25 ms. Throughput: 70.94 iter/sec.

Notice how 4 workers is not getting more work (iter / sec) done than 3 workers.

With 2 DIMM in 2 channels, the system uses 80 W and gets these benchmark results.

Timings for 4096K FFT length (1 cpu, 1 worker): 24.09 ms. Throughput: 41.51 iter/sec.
Timings for 4096K FFT length (2 cpus, 2 workers): 25.11, 24.77 ms. Throughput: 80.20 iter/sec.
Timings for 4096K FFT length (3 cpus, 3 workers): 27.38, 27.38, 27.50 ms. Throughput: 109.42 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 32.22, 32.11, 32.09, 32.08 ms. Throughput: 124.51 iter/sec.

If the CPU was the bottleneck, we would expect 4 worker to get 4x the
throughput of 1 worker, or 4x 41.51 = 166 iter / sec instead of the 124.5
I am getting.

The 4096K FFT is used for exponents around 70,000,000 .
Doing other types of work, like double check, trial factor, or P-1 factoring
may get different results.

Hope this helps.

[Madpoo]

Originally Posted by CRGreathouse

I'm interested in comparing CPUs in terms of (among other things!) power consumption. How reasonable is it to assume that power consumption at full load is equal to TDP?

If your cooling is up to the task, you will get pretty close to TDP specs. Close enough that, from an engineering standpoint, you might just say it's the same.

In practice, if you're pushing the thermal limits, your CPU probably won't run at the fastest turbo speed (c0 or whatever) so it'll clock down just enough to keep within the "thermal" part of TDP, which might mean it's power usage is less than the max.

I've mentioned before that on my dual-socket HP Proliants, when stress testing I'll see CPU #1 run at full speed (full turbo state), but CPU #2 is one tick slower, because it's running a bit hotter. I think it has to do with air flow through the system, like air being drawn over the 15K RPM disks up front, design of the air baffle, the proximity of CPU #2 to the array controller (which runs really hot), etc.

Still, you'd be within spittin' distance of max power, like within a couple watts.

[VictordeHolland]

You could get a 15$/£/€ kill-a-watt meter and measure the power from the wallsocket :).
My 2500K @4.0GHz (95W TDP) system running P95 is consuming 120W at the wall. Overclocking not being good for perf/Watt.
From that 120W, let's say 10% (12W) is wasted in the PSU as conversion losses, 10W for HDD and casefans and another 10W for the motherboard. That leaves about 90W for the CPU, close to the TDP of that CPU.

[CRGreathouse]

Originally Posted by VictordeHolland

You could get a 15$/£/€ kill-a-watt meter and measure the power from the wallsocket :).

Useful if I want to measure my current system, not so useful when I want to compare chips I'm considering buying.

[kladner]

Don't these results give at least some idea of how close to TDP real usage might come? Of course, measuring at the mains, you have to take into account the efficiency of the PSU, and such things as number and power consumption of drives and fans. Victor's assignment of 10 w for drive and fans seems a little conservative to me. But then, I have a bunch of spinning things in there, both drives and fans.

It still seems that such measurements can make for more informed estimates, especially if you can narrow down which parts are using what power,

[Madpoo]

Originally Posted by kladner

Don't these results give at least some idea of how close to TDP real usage might come? Of course, measuring at the mains, you have to take into account the efficiency of the PSU, and such things as number and power consumption of drives and fans. Victor's assignment of 10 w for drive and fans seems a little conservative to me. But then, I have a bunch of spinning things in there, both drives and fans.

It still seems that such measurements can make for more informed estimates, especially if you can narrow down which parts are using what power,

Don't forget the power used by the memory as well. It can add up. What's typical? Like maybe 3 - 5 W per module?

[kladner]

Originally Posted by Madpoo

Don't forget the power used by the memory as well. It can add up. What's typical? Like maybe 3 - 5 W per module?

Yes. It's all part of the energy budget.

[VictordeHolland]

I don't know if the power usage in Core Temp 1.0 RC6 is accurate (it's reading 89.4W) , as the voltage reported is off too compared to CPU-Z (which I would thrust more).

How I estimated the 10W for fans and harddisk:
SSD and HDD (3W),
2x 120mm (rated 12V 0.2A) fans = 4.8W
CPU fan maybe 2-3W

The 10W for the motherboard/memory is just a guess, but it could be more as 2 memory modules might already consume that as pointed out by madpoo.

[kladner]

Originally Posted by VictordeHolland

I don't know if the power usage in Core Temp 1.0 RC6 is accurate (it's reading 89.4W) , as the voltage reported is off too compared to CPU-Z (which I would thrust more).

How I estimated the 10W for fans and harddisk:
SSD and HDD (3W),
2x 120mm (rated 12V 0.2A) fans = 4.8W
CPU fan maybe 2-3W

The 10W for the motherboard/memory is just a guess, but it could be more as 2 memory modules might already consume that as pointed out by madpoo.

I withdraw my skepticism. That all sounds reasonable. As I said, my system is populated with a number of HDDs and some rather powerful fans. It skews my perceptions.