GPU Work Units 13416 - Nvidia 2080TI not being fully used

[WinstonCooper]

Hello!

I'm a first time poster - but just wanted to see if anyone else has noticed that at least some (seems to be most) of the work units which are numbered 13416 are not fully utilizing 2080 GPUs.

I am running the latest NVidia drivers, and most WUs put me at or close to 4 million PPD on my 2080ti EVGA Hybrid card. This particular WU runs at 1.8 - 1.9m PPD.

Lately, I have been getting quite a few of these #13416 WUs. The PPD is lower but so is the GPU utilization.

Normally my card runs at 99% all day, at 53-55c due to the AIO for cooling. With these WUs I don't exceed 65% utilization or even approach the cards maximum power draw.

Any ideas as to what's going on? Most WUs are fine - it appears to just be this this set.

I tried to include a picture and a thumbnail but apparently there are some rules in place that prevent me from doing so. I am unable to assist with providing visuals to assist in troubleshooting.

[JimboPalmer]

Welcome to Folding@Home!

Traditionally, GPU are divided by capability, not performance. So my GTX 1650 (896 shaders) and your RTX 2080 ti (4352 shaders) would get the same Work Units, as both are CUDA 7.5; but a GTX 980 might not. (2048 shaders, but the older CUDA 5.2)

Lately F@H has been trying to do more performance matching as High End GPUs have more shaders than the proteins have atoms. You want to give those smaller proteins to wimpier cards like mine, and run massive molecules on beefy cards like yours.

So far that is not perfect. And as i say, over 10 years of tradition need to be coded around, so it may take a while.

If you want the COVID-19 solved rapidly, and many do, then some GPUs will be assigned smaller proteins that would be less than ideal from a Points Per Day perspective.

[WinstonCooper]

That makes perfect sense. I suspected something as such. I wonder if it would be effective to attempt to add a second GPU slot ( or run multiple FAH instances on virtualized OS’s) if these kinds of work units persist? I don’t know if that’s feasible, I guess it’s potentially possible with a hypervisor like ESXi that’s capable of GPU hardware level access - maybe....

I doubt that juice would be worth the squeeze though, since most WUs aside from these will peak the card.

Thanks for the info!

[JimboPalmer]

if you think about this from AMD and Nvidia's point of view, the vast majority of GPU buyers are gamers, who want very little overhead to slow FPS (Frames Per Second), so a Driver with the ability to run multiple projects at once is not really going to sell cards. A simple driver that allocates everything to one task makes gamers happy. You and I wish for more, but we just don't drive sales.

[BobWilliams757]

Also keep in mind that with powerful cards, there are times when the WU just won't fully utilize it. Even with my meager onboard graphics (Vega 11) that would make JimboPalmer's GTX1650 seem like a powerhouse, I on occasion get work units that won't fully utilize the card. In my case it's usually the units with very low atom counts, and while my onboard card often gets the higher PPD from such units, people with powerful cards find that they end up with a low (to them) points return.

13416 is also one of the projects that shows a lot of variation run to run, and one they are studying to gain info about how to set up future projects. From my understanding it's part of the Covid Moonshot project and as such a very high priority. So even though they are learning and working out bugs as we go, it's for good reason.

[JohnChodera]

Thanks for checking in! Not all projects reach the system size that is needed to make maximal use of an RTX 2080 Ti. Project 13416 is a project supporting the COVID Moonshot (http://postera.ai/covid) where we really need to grind through as many WUs as we can for a medium-size (~80K atom) system to help prioritize which molecules the chemists should synthesize. We're also working to model the SARS-CoV-2 main viral protease dimer, which should push the RTX 2080 Ti harder, as well as investigate the peculiar mechanism by which the Vir antibody S309 (now moving to clinical trials) is able to neutralize the virus so we can better understand resistance (which will involve much larger systems). When we get to that stage in a few days, we'll try to rebalance the project constraints to better take advantage of faster GPUs like the 2080 Ti.

~ John Chodera // MSKCC

[aetch]

It's probably worth noting that 13416 has been a very rocky project for the load it puts on systems.
Personally, on my RTX 2070 SUPER and GTX 1080 Ti I've seen TPF ranging from 1min 30 seconds to 3miins 30 seconds with GPU load ranging from 60% to 95%+.

[markdotgooley]

13416 units usually use either of my RTX 2060 cards pretty well. A 2080ti just might be overkill (some tables of GPU performance note that some WUs cannot use a 2080ti effectively). Maybe with the new generation of GPUs, GPU abilities and WUs will be matched better. It’d be rather disappointing for me to invest in faster new cards only to find they do about as well as my current ones.

[Skajaquada]

I can say my multiple RX5700 XT work very well with the 13416. Yes sometimes it has less performance, but not so much as intended.
The PPD Range from the 13416 Projects goes from 70% to 130% of the average PPD from 1.35 million.
So "bad" 13416 runs are not a real problem on this GPU.