4070ti low power usage...

[Mxyzptlk]

Curious, just installed a 4070ti and I use MSI Afterburner to monitor the power usage as well as using MSI to power limit it too.

I have no power limit active, but the card is only using ~70% (200 watts) of the power rating for the card (285 watts). I have no PPD data to compare this card too and have only run 2 WU's on it so far... averaging ~8.5MPPD. The first WU was 18448 and the second WU is 18449.

I have run a quite few other cards and have never noticed this behavior before. Usually a card with no power limit runs at 100% of power while folding. Typically I power limit at around 70-80% to save some power without sacrificing a bunch of PPD for sciencing...


(I also use EVGA Precision X1 and confirmed the same that I found in Afterburner)

[Vester]

If you look at Task Manager | Performance | GPU0, you will find that FAH does not utilize 100% of the GPU (in the 3D section).

[Mxyzptlk]

I am not looking a Task Manager. Task Manager does not show the true GPU usage.

Now that I have had it for a few days, I am coming to the conclusion that it may just be related to the individual WU.. but it is still very odd behavior.

[Vester]

My GPU utilization is about 70%. You can see the GPU utilization in HWMonitor, too. Some members run two installations of FAH in different directories to fully utilize the GPU. I have not done that.

[toTOW]

I have run a quite few other cards and have never noticed this behavior before. Usually a card with no power limit runs at 100% of power while folding. Typically I power limit at around 70-80% to save some power without sacrificing a bunch of PPD for sciencing...

I have a 1070 that I've never seen hit its power limit, no matter what I run on it ... on the opposite, I have a 980 that is regularly going beyond 100%, even when folding ... I don't think you can find a general behaviour here.

I should have my own numbers from a 4070 Ti by the end of the month, I'll let you know what I see ...

Some members run two installations of FAH in different directories to fully utilize the GPU. I have not done that.

I don't know where you saw this false information, but this is completely useless and usually counter productive !

[Mxyzptlk]

Interesting.

Thanks all!

[FaaR]

I have run a quite few other cards and have never noticed this behavior before. Usually a card with no power limit runs at 100% of power while folding.

Power useage behavior is not an exact science, this can/does vary between different pieces of hardware and the particular loads you run on them. As FaH is purely compute oriented, it doesn't exercise many parts of a traditional 3D accelerator (like for instance the geometry setup bits, texture fetching/filtering bits, polygon rasterization bits etc); these parts of the GPU would otherwise consume a sizeable portion of the chip's power budget, so it could be that your board has a power limit which is set so generously that it allows both shader arrays and polygon pipe-related hardware to run at high performance levels.

Thus, if you run purely compute, your power draw would not hit the max ceiling (as you have a lot of GPU hardware idling.)

It could also be that the particular work units you have received (a whopping two isn't really a sufficient statistical data set! don't max out your shader array for one reason or another, and that's why your power useage is looking on the low side.

Regardless - why worry? Be happy! You're using less power, which is good!

[mew905]

So my 4070ti typically gets about 215w power draw while folding and maintaining a relative 110-14 million PPD. However I noticed a massive PPD drop for 18717 WUs. Like 5.3M PPD (not down 5.3M, down TO 5.3M).

Checking power draw in GPU-Z I'm seeing ~140w power draw typically. GPU usage still sits around 83% (fairly normal), so I've been trying to figure out how to get that power draw back up to normal.

Now, I would have accepted this as standard WU deviation, but my laptop's 3070 on the same WU, drops from ~3M PPD to 2.2M PPD, a 27% drop, as compared to my 4070ti's 61% drop. My 4070ti has more PCIe bandwidth than the 3070 so the issue isn't the bandwidth. Is it the CPU? (8 core Xeon E5-2667 V2 with 4070ti vs Ryzen 5800H with the 3070m)

[Mxyzptlk]

18717 has a low atom count and those type of WU's just do not play well with 30XX and even worse with 40XX Cards.

[mew905]

18717 has a low atom count and those type of WU's just do not play well with 30XX and even worse with 40XX Cards.

18717 has at 77k atoms. Meanwhile 18715 has 100k atoms and provides much higher PPD than 18601 (334k). I imagine there's far more to it than just raw atom count that adds to a WU's complexity.

That said, switching to v8 client, I'm seeing 6.3M PPD now. IIRC V8 is designed to reduce some CPU overhead (and kill the API I use to monitor clients... yaay...) which again points me toward my CPU not being fast enough. Despite having 8c16t, it only seems to manage to be able to boost up to 3.8GHz.

[toTOW]

IIRC V8 is designed to reduce some CPU overhead (and kill the API I use to monitor clients... yaay...) which again points me toward my CPU not being fast enough.

v7 and v8 use the same compute cores, which don't change when you change client version ...

[BobWilliams757]

18717 has a low atom count and those type of WU's just do not play well with 30XX and even worse with 40XX Cards.

18717 has at 77k atoms. Meanwhile 18715 has 100k atoms and provides much higher PPD than 18601 (334k). I imagine there's far more to it than just raw atom count that adds to a WU's complexity.

That said, switching to v8 client, I'm seeing 6.3M PPD now. IIRC V8 is designed to reduce some CPU overhead (and kill the API I use to monitor clients... yaay...) which again points me toward my CPU not being fast enough. Despite having 8c16t, it only seems to manage to be able to boost up to 3.8GHz.

I agree that there is more than atom count involved in project variation, but there is no doubt that atom count vs shaders and associated hardware do have a direct impact. As with your examples of different projects on your 4070 vs 3070 GPU's variation might be even less on a 2 series card. I think the entire picture of cores/shaders/frequencies/memory, etc all come into play to determine the "sweet spot" where any certain GPU really shines, or at times struggles. When I folded on my 2400G iGPU it thrived on very low atom count projects, at say 5000 and below it would quickly close the performance gap to near GPU's that were easily 5-8 times more powerful. They were in a weak spot, the iGPU was in a sweet spot.

As for the CPU impact, as with many I don't think that is the issue you are experiencing. Since I don't have a quick GPU I can't confirm or deny that but you easily could. Do some testing and during that testing pause the project, reset some CPU settings (lower clocks, turn off CPB or such) and then resume. Watching TPF's is not going to take long when such a fast card can process several percent and reach stability fairly quickly.

[jchang6]

I am getting frequent assignment of 18717 recently, or at least I am noticing it now. On 3080 Ti, I usually get 8M PPD? (lar says 6M). On the 2080 Super, I usually see 3M. For project 18717, I am seeing 6M PPD on the 3080 Ti and 1.5-1.8M on the 2080 Super. It seems to be steady on the 3080Ti and fluctuates on the 2080 Super, more in the low direction. All my 200 watt+ cards are reduced to 80% power via MSI afterburner.
i.e., the impact of 18717 seems to be greater on the 2080 Super than on the 3080 Ti

[BobWilliams757]

I don't think anyone is thrilled with the points returns on 18717. Often the ones that the really high shader count GPUs don't like run fairly quick and cards with less shaders but quick clocks and decent memory.

The variance of 18717 on my 1660 Super doesn't take it to the slowest projects, but it's far from a top performer as well. At 75% power it gives about 1M PPD, at the same power a number of others can get 1.3-1.5M. But it seems to me that often the higher points projects seem to pay out more points to most hardware, so at least its' somewhat consistent. There are trends in atom counts, but always exceptions.

Benchmarking alone has to leave some variance. They can only spend so much time and get it close, otherwise we would all get almost consistent PPD returns... which might be kind of boring really.