RTX 4090....underperforming?

[emperor886]

Howdy folks,

I've been folding for years and never really had an issue, I generally just let the GPU run whenever I don't need it any never pay any attention to PPD. Recently though (3.5 weeks) I got the 4090, and having heard about the mind-boggling numbers that people were getting, I started watching my PPD...but was disappointed. I'm getting around 9.7-11.5m PPD on full. I'll link my donor page below (idk if that's helpful or not) but I just realized today that I had a typo in my team number, so not sure if the points for those WUs went into the void or default.

- CPU also runs, research set to high priority (maybe getting harder WUs based on having better hardware?)
- OC curve set by Afterburner scanner, power limit 133%, temp limit 88C (never goes above 59C)
- Latest studio driver and CUDA version
- According to Windows and GPUz, near full GPU utilization
- Have tried DDU & fresh driver install
- Have also tested at default clock & limit settings
- If it matters: Aorus X570 + Ryzen 9 3900X + 64GB DDR4 3600Mhz

https://stats.foldingathome.org/donor/id/101298
https://imgur.com/a/lvhBPbJ

These numbers are still significantly higher than my previous card so if they don't go up, it's still a win overall. But if I can start getting numbers near what others are achieving, that would be ideal.

[Mxyzptlk]

Several things to look at:
- Are you getting a lot of Cancer WU's? Those have pretty low PPD production of 4000 series cards (3000 series too). However, I haven't seen many Cancer WU's lately...
- Having the CPU fold at the same time will lower the PPD on the GPU. Not as much as yours is showing, but it will be lower.
- When Folding, what is the actual GPU power draw?

I don't know too many Folders that run their GPU's at 100%+ power level. I run all of mine in the 65 - 80 % range. At 80 I have never seen much of a PPD loss if any.

Not sure If I helped, but I'm very curious too why your doing roughly 50% of the normal PPD....??

[PaulTV]

If you run discord app, switch off GPU acceleration. Also see https://www.tomshardware.com/news/disco ... ock-speeds

[kiore]

First thing is that people tend to report their best numbers under ideal conditions, so yours may rarely meet those expectations. However follow the suggestions above and I would add letting your CPU units complete by set to finish and see how your GPU does by itself, use GPUz or similar to see actual utilization. There will be variation sometimes extreme ones with very high end cards between units so run a few different ones before deciding what you GPU is doing by itself. I suspect that the wrong team name if you are using a passkey (and if you aren't then do so) will be credited to you generally but not on the team name, this however will not impact on the projected PPD showing on your side. The first few frames on a new unit may be wildly out so don't count your PPD until you have a reasonable number completed, I usually discount any projections below 4 frames but by 10 you will have a reasonable projection.
All the best.

[bikeaddict]

See this post from LAR Systems.

https://linustechtips.com/topic/902973- ... t=15776205

On 1/30/2023 at 8:46 AM, LAR_Systems said:
"What I commonly see with new GPUs is that F@H will toss work units that can be too easy or more less not optimized for the new GPU and it results is crazy results for the first few weeks.

"Then things seem to settle (average) down in almost all cases which suggests they benchmark then put the GPUs to work on the kind of work where they can get the most out of the hardware."

[emperor886]

Several things to look at:
- Are you getting a lot of Cancer WU's? Those have pretty low PPD production of 4000 series cards (3000 series too). However, I haven't seen many Cancer WU's lately...
- Having the CPU fold at the same time will lower the PPD on the GPU. Not as much as yours is showing, but it will be lower.
- When Folding, what is the actual GPU power draw?

1. No idea, I'm currently working on project 18449 which says Influenza, idk how to check history though (if it's even possible). To be clear, I don't care in the slightest about points, just contributing to the cause. And I assume that points are a fairly decent indicator of how much work is being done (or how efficiently) and I just want to maximize that.

2. Set my CPU to finish so I don't waste the WU, it'll be done tomorrow morning and I'll monitor for 7 days after that before updating here again.

3. So the previous screen I posted from GPUz showed 16-pin power at 374W and GPU chip at 332W, with a load of 91%. I just restarted and began monitoring immediately, after 10 minutes I'm seeing 16-pin at 255W, GPU chip at 219W, and load at 79% (all averages). Power consumption shows an average of 59% TDP. I had it save to a log file: https://pastebin.com/yZfXBDRj

If you run discord app, switch off GPU acceleration. Also see https://www.tomshardware.com/news/disco ... ock-speeds

Done and restarted. Thank you!

First thing is that people tend to report their best numbers under ideal conditions, so yours may rarely meet those expectations. However follow the suggestions above and I would add letting your CPU units complete by set to finish and see how your GPU does by itself, use GPUz or similar to see actual utilization. There will be variation sometimes extreme ones with very high end cards between units so run a few different ones before deciding what you GPU is doing by itself. I suspect that the wrong team name if you are using a passkey (and if you aren't then do so) will be credited to you generally but not on the team name, this however will not impact on the projected PPD showing on your side. The first few frames on a new unit may be wildly out so don't count your PPD until you have a reasonable number completed, I usually discount any projections below 4 frames but by 10 you will have a reasonable projection.
All the best.

My CPU unit will complete tomorrow and then I'll monitor for 7 days and report back (if everything here doesn't improve the numbers).

See this post from LAR Systems.

https://linustechtips.com/topic/902973- ... t=15776205

On 1/30/2023 at 8:46 AM, LAR_Systems said:
"What I commonly see with new GPUs is that F@H will toss work units that can be too easy or more less not optimized for the new GPU and it results is crazy results for the first few weeks.

"Then things seem to settle (average) down in almost all cases which suggests they benchmark then put the GPUs to work on the kind of work where they can get the most out of the hardware."

Ah, okay! This makes sense.

And in a surprising twist, literally as I was finishing this post I noticed my estimated PPD had hit 23.2m. It is now working on project 18213 (Alzheimers). Hopefully I continue to be sent WUs like this can use my card to its full potential. GPUz shows the same power and utilization as I mentioned above, so this WU seems to be more optimized for my hardware (the one that was running just prior to me starting this reply, 18449 Influenza, was showing 12.1m).

[Mxyzptlk]

The cards do there best on Alzheimers, influenza not so well... You can set a preference for WU's in the config.

[Mxyzptlk]

See this post from LAR Systems.

https://linustechtips.com/topic/902973- ... t=15776205

On 1/30/2023 at 8:46 AM, LAR_Systems said:
"What I commonly see with new GPUs is that F@H will toss work units that can be too easy or more less not optimized for the new GPU and it results is crazy results for the first few weeks.

"Then things seem to settle (average) down in almost all cases which suggests they benchmark then put the GPUs to work on the kind of work where they can get the most out of the hardware."

That was in response to a question from me about the 4070ti. The 4090 and 4080 have a lot of WU's and are pretty settled in.

[BobWilliams757]

It's always somewhat the luck of the draw with the work assignment servers, and with logging it's easy to see trends.

With the newer cards having more cores, generally speaking low atom counts will result in lower PPD, higher atom counts take advantage of the hardware more and produce higher PPD. Even within those ranges variance in base point assignments, project time simulated, etc all impact the scores.

Even if a person isn't concerned with PPD using HFM to monitor results, along with something like GPU-Z to keep track of average power use, allows tuning to best efficiency, best PPD, or a compromise between the two at times. It also shows the trends, and at time using one preference over another might be favorable for max productivity.

It seems like as GPUs get more and more powerful, it's already reached a point where the variance is going to be greater overall. There just aren't currently enough of the larger atom count projects to keep all the power cards busy.

[thiblec71]

Do you see a big difference of running Win 10 or Win 11?

[thiblec71]

I am currently folding some influenza projects at 3mio PPD with my RTX 4090 and i5 13600k on windows 10.
GPU load around 75% and power consumption at 50-52%

I have the latest nvidia gefore studio driver.

What am I missing to fold faster?

[Lazvon]

Just installed my 4090. Briefly this afternoon 4090 and 4080 had some project ended in 424 or something... 20M estimated PPD on the 4080 and 24M estimated on the 4090. Now back to 18448 and 18449... only 4090/10.7M and 4080/8.9M. It is SOOOO much quieter than the 3090Ti was though. Wow.

[thiblec71]

Now back to 18448 and 18449... only 4090/10.7M

@Lazvon, I get the same fold rate on those projects but the GPU load is at 75% and wattage at around 250w.
Does your 4090 run higher?
Are you on Win, if so 10 or 11?

[Lazvon]

I haven’t checked wattage or utilization. But I did notice it doesn’t crack 60 degrees on the 4090 (Gigabyte Windforce V2). Just used MSI Afterburner and leveraging the curve it set.

Win10 until they pry it from my cold dead hands… work just gave me a Win11 VDI … and all four kids just got new “gaming rigs” with Win11… doesn’t look too bad.

[toTOW]

Just installed my 4090. Briefly this afternoon 4090 and 4080 had some project ended in 424 or something... 20M estimated PPD on the 4080 and 24M estimated on the 4090. Now back to 18448 and 18449... only 4090/10.7M and 4080/8.9M. It is SOOOO much quieter than the 3090Ti was though. Wow.

p12424 seems to be doing pretty well on bigger GPUs