More Power Efficient To Run GPUs in x2 Mode?

JohnSmith2 · January 15

I render a lot of 3D animations with multi GPU setups and I have experimented a lot to optimize for power efficiency rather than render times. I've found that power limiting the GPUs to 50% results in a pretty large efficiency gain (more frames per watt hour) but one thing I didn't expect to make a large difference was how many PCIe lanes the GPU is getting.

Because I'm using a consumer platform (AMD 7950X, X670 motherboard), the motherboard I use doesn't have very many PCIe lanes. I was experimenting with the GPUs in different slots to make room for a network card and this is what I observed:

PCIe Setup Lanes

(1x EVGA SC2 GTX 1080 Ti)

Redshift Vulture Benchmark

Render Time (Seconds)

Average GPU Power Draw

(watts) measured by GPU-Z

x16

733

122.4 W

x4

755

116.5 W

x2

790

81.1 W

(I only used one GPU for these benchmarks, and it was the same GPU, just moved to a different slot on the motherboard. All PCIe slots tested are PCIe 4.0)

From x16 to x4, the differences are pretty small, but I was shocked to see the difference from x4 to x2; 35 watts fewer but only a 4.6% reduction in performance. Not sure if you would notice something similar for newer GPUs or if this is architecture dependent.

Of course whole system power draw is ultimately what matters but even when I add in the consumption of the other components and measure the power draw from the wall, this is still an efficiency gain of 2% to 4%. Not huge, but unintuitive to me, especially considering there is an efficiency loss of about 1% to 2% from x16 to x4 mode.

Any idea why the power draw is so much less when running in x2 mode, considering the x4 power draw is so similar to x16?

jaslion · January 15

Gpu isn't being fed enough data so it ends up lowering it's clock speed as it doesn't need to work so hard.

JohnSmith2 · January 15

4 minutes ago, jaslion said:

Gpu isn't being fed enough data so it ends up lowering it's clock speed as it doesn't need to work so hard.

Interesting. So this implies I could probably achieve the same efficiency in an x16 or x4 slot if I artificially restrict the clock speeds. I should have recorded those while testing.

Bismut · January 15

2 minutes ago, JohnSmith2 said:

Interesting. So this implies I could probably emulate the same efficiency in an x16 slot if I artificially restrict the clock speeds. I should have recorded those while testing.

Yup, could be. I would also recommend looking into undervolting, my R5 1600AF uses about 20W less than before, while actually performing better (which I attribute to the lower thermals).

jaslion · January 15

31 minutes ago, JohnSmith2 said:

Interesting. So this implies I could probably achieve the same efficiency in an x16 or x4 slot if I artificially restrict the clock speeds. I should have recorded those while testing.

Do keep in mind that this can severly effect performance depending on render load.

Id honestly keep it at like x8 and just lower power limits. Then if you need a quick quick render you can let em go full tilt if needed.

manikyath · January 15

59 minutes ago, JohnSmith2 said:

Interesting. So this implies I could probably achieve the same efficiency in an x16 or x4 slot if I artificially restrict the clock speeds. I should have recorded those while testing.

as a sanity check, run it with x16 but power limit the GPU (refer to people who actually fiddle with OC for how to set power limits)

i'd assume that limiting to 85 watts should roughly match the results for x2.