Multigpu power consumption numbers

[ehhhhhhh]

Recently there were a few questions regarding to the power consumption of multigpu setups so I thought I share some numbers.

<-- System is on the left

Firstly, some details on the environment:

The CPU runs at 3.8Ghz with CnQ disabled @ 1.45V and the HT runs at 2.2Ghz
The motherboard voltages have been raised by 0.1V.
The memory voltage is upped by 0.1V too

The gpus are unlocked 6950 running 2x 1080p displays.

All the readings have been carried out at the socket and consist only the PC.


Idle power consumption: 210W

Prime95 (blend test on all cores): 340W

Arma2 (very high details, 1080p, no CF support): 330W

Metro 2033 (very high, 1080p): 500W

3Dmark 2011 combined GPU+CPU test: 450W

Hope this helps choosing the right PSU

[c.ruel]

I have always wondered about this because on AMD's site for nearly all of their previous generation cards it said you needed 600w for an xfire setup. Yet on the review sites it was always like "850w+ psu needed". It is hard to know who to believe.

The fact that gpu manufactures do not specify how many amps you need as well as watts overall just confuses the picture even more.

I know your numbers were just done with a wall meter so won't be incredibly reliable but still very interesting to know.

[oolon]

It also does not help when manufactures cannot specify a power level correctly. I remember having an argument with my manager about this, we had HP proliant servers and they stated in the documentation an 8 amp current draw. My manager through didn't believe me when I told him that it was wrong, until I pointed out a rack of proliant servers and said they all used the same 16A supply (strictly 32a (2x16A)) and if we went by the specs no way could that power the whole rack of 8 servers.

[ik9000]

To get best efficiency out of a power supply you generally want to be running it at about 50% -75% of its rated power. So for a 500W draw it's not unreasonable to suggest a 650W-1000W PSU, though the bigger the rating, the more expensive the PSU, and generally, though not always the noisier it can be when running (above 50% load).

Let's look at 80 plus Gold efficiency for a PSU on 230V 50Hz AC. Requirements are min 92% at 50% load and 88% at 20% and 100% load

1000W PSU, supplying 500W, 92% efficient = 543W being consumed from the mains

500W PSU, supplying 500W, 88% efficient = 568W being consumed from the mains

650W supplying 500W (77% load), efficiency min 90% (approx) = 555W being consumed from the mains.

These are min values for a gold rated PSU. Looking at the charts for individual PSUs however, many exceed these min efficiencies, and some offer 91-92% efficiency from 45-80% load so you can get away with using a lower rated (and less expensive) PSU without getting a more expensive electricity bill for it.

If you get a 650W PSU offering 92% efficiency at 77% load then that is a better deal than a 1000W PSU offering 92% efficiency at 50% load if you only need 500W. It will probably be quieter too, though not all manufacturers are equal when it comes to running noise.

[Domestic_Ginger]

I assume this is measured at the wall; whats the efficiency.

I would suggest running prime and either furmark or other intensive GPU program in order to get absoloute power draw.

Furmark and the 5k series did not get on so I'm no sure how the 69**s survive. Whats teh final voltage on your ram as a 1V bump sounds alot.

[ik9000]

Going from the OP system spec he's using a Corsair TX750w PSU,

http://www.corsair.com/tx750w.html



500W measured as being drawn from the mains with 86.5% efficiency and you hit roughly 433W actually being supplied by the PSU to the system.

[ehhhhhhh]

I assume this is measured at the wall; whats the efficiency.

I would suggest running prime and either furmark or other intensive GPU program in order to get absoloute power draw.

Furmark and the 5k series did not get on so I'm no sure how the 69**s survive. Whats teh final voltage on your ram as a 1V bump sounds alot.

ahh the ram meant to be +0.1..i'll correct it in a minute

Considering real world scenarios it's not likely that anything maxes out the CPU and the GPU too. though I'm happy to run the required tests.

having said that I'll need to find something more demanding than furmark as it won't put 100% load on both cards even on maxed out (no throttling occurs in terms of frequencies).

The toughest thing is metro2033 so far

I know your numbers were just done with a wall meter so won't be incredibly reliable but still very interesting to know.

I have 2 meters giving the same values, also just for testing I've connected things with known powerdraws (heater, hover, etc) and it seems that it's very accurate depending on those numbers.
eg. an 1800W heater fluctuates between 1790-1810....

mind you I'm not an expert of electricity and stuff