The PPA-reset trick worked - I have a new task running now. Another satisfied customer.

The completion estimate at 10% was 43.5 hours, both by extrapolation and by setting <fraction_done_exact/> in app_config.xml

It's a ADRIA_New_KIXcMyb_HIP_AdaptiveBandit task. I ran a couple of these about 10 days ago, under the old app: they took about 33 hours - previous 'D3RBandit_batch*' tasks had been 28 hours on average. Cards are GTX 1660 super.

So there's a possibility that the new app is slower, at least on 'before cutting edge' cards.

Some of the tasks were even longer. I have two more 2080ti reports that were 55,583 and 56,560s respectively. On Identical GPUs running the same clocks. There seems to be Some variability. If you use the slower one it puts it closer to 30%. This exposes the flaw of using a single sample to form a conclusion. More data is required.

Also note that I’ve been experiencing performance issues with this specific card. I believe it’s underperforming due to some incorrect power limiting behavior (I’ve done a lot of load testing and cross referencing benchmark results with others online). I have a replacement on the way to test.

These ADRIA tasks have hard coded reward. It isn’t necessarily based on run time. They increased the reward from the D3RBandit to these KIXcMyb tasks, but since they stopped distributing the CUDA 10 app, we can’t know for sure if the tasks are just longer or if there’s some inefficiency in the new 11.2 app that’s slowing it down. If the tasks aren’t longer, then the new app is almost 30% slower than the old CUDA app

---

I looked back into the 'job_log_www.gpugrid.net.txt' in the BOINC data folder to get my comparison times. I haven't run many AdaptiveBandits yet, but I think the 'D3RBandit_batch*' time was a robust average over the many sub-types.

After cross referencing runtimes for various Windows hosts, I think the new app is just slower. Windows hosts haven’t experienced an app change (yet) and haven’t shown any sudden or recent change in run time with the KIX AdaptiveBandit jobs. This suggests that that tasks haven’t really changed, leading the only other cause of the longer run time to be a slower 11.2 app.

I also noticed that the package distribution is different between the CUDA 10 and 11.2 apps. 10 included some library files that are not included with 11.2 (like cudart and cudafft libraries) so the app may have been compiled in a different way.

I hope Toni can bring the app back to par. It really shouldn’t be that much slower.

---

GTX 1080
# Speed: average 75.81 ns/day, current 75.71 ns/day

RTX 2070S
# Speed: average 134.99 ns/day, current 132.17 ns/day

RTX 3070
# Speed: average 159.15 ns/day, current 155.75 ns/day
https://www.gpugrid.net/result.php?resultid=32632515
https://www.gpugrid.net/result.php?resultid=32632513

only task yet is with 3070 and ended after 18-19 hours
3000-series looks slow with 11.2 but they works. Progressbar and estimate looks to be close expected time and 2070 could probably end after around 21 hours.

It would be great if Toni could make application to print from progress.log

Had to add PPA for libboost needed and did try update one host to 21.04 to get latest boost but did not work.

Where did you get the ns/day numbers from?

But it’s not just 3000-series being slow. All cards seem to be proportionally slower with 11.2 vs 10.0, by about 30%

---

Go to slot folder and cat progress.log

Yes looks like all cards are affected on new application. I compare with 1000-series also but do not have numbers of ns/day for them.

Where did you get 469 driver? Can't see it on nvidia site or PPA.

It’s not real. I’ve manipulated the coproc_info file to report what I want.

Actual driver in use is 460.84

---

RTX 3070
# Speed: average 159.15 ns/day, current 155.75 ns/day
https://www.gpugrid.net/result.php?resultid=32632515
https://www.gpugrid.net/result.php?resultid=32632513

only task yet is with 3070 and ended after 18-19 hours
3000-series looks slow with 11.2 but they works. Progressbar and estimate looks to be close expected time and 2070 could probably end after around 21 hours.

The WU you linked had one wingman run it as cuda 10.1 and the other as 11.21 with 155,037 seconds versus 68,000. Isn't that faster?
https://www.gpugrid.net/workunit.php?wuid=27075862
What does ns mean?

I didn’t use the time remaining estimate from BOINC. I estimated it myself based on % complete and elapsed time, assuming a linear completion rate.

I usually employ the same method, since Progress % shown by BOINC Manager is quite linear.
At my low-end GPUs, I'm still waiting for the first task to complete :-)
Evaluating the small sample of tasks that I've received, tasks for this new version are taking longer to complete than previous ones (lets say "by the moment")
Estimated completion times for the 5 GPUs that I'm monitoring are as follows:



The last three GPUs are Turing GTX 1650 ones, but different graphics cards models and clock frequencies.
An editable version of the spreadsheet used can be downloaded from this link

RTX 3070
# Speed: average 159.15 ns/day, current 155.75 ns/day
https://www.gpugrid.net/result.php?resultid=32632515
https://www.gpugrid.net/result.php?resultid=32632513

only task yet is with 3070 and ended after 18-19 hours
3000-series looks slow with 11.2 but they works. Progressbar and estimate looks to be close expected time and 2070 could probably end after around 21 hours.

The WU you linked had one wingman run it as cuda 10.1 and the other as 11.21 with 155,037 seconds versus 68,000. Isn't that faster?
https://www.gpugrid.net/workunit.php?wuid=27075862
What does ns mean?

nanosecond
https://en.wikipedia.org/wiki/Nanosecond#:~:text=A%20nanosecond%20(ns)%20is%20an,or%201%E2%81%841000%20microsecond.


Yes there big gap to runtime on other host but it was also using NVIDIA GeForce GTX 1070

It’s not real. I’ve manipulated the coproc_info file to report what I want.

Actual driver in use is 460.84

Ok why i ask was that device name is unknown for my 3080Ti and had some hope that driver you used would fix that.
So i could go coproc file and edit instead.

It’s not real. I’ve manipulated the coproc_info file to report what I want.

Actual driver in use is 460.84

Ok why i ask was that device name is unknown for my 3080Ti and had some hope that driver you used would fix that.
So i could go coproc file and edit instead.

What driver are you using? The 3080ti won’t be detected until driver 460.84. Anything older will not know what GPU that is.

---

Greger, I just can't get my head around what it means. So out of the 8.64E13 ns in a day you only calculate for 159 ns??? I'm not familiar with that figure of merit.

BTW, my 3080 is running 465.31. Still waiting to catch a WU after the PPA, reboot & reset.

The nanoseconds will be the biochemical reaction time that we're modelling - very, very, slowly - in a digital simulation.

Greger, I just can't get my head around what it means. So out of the 8.64E13 ns in a day you only calculate for 159 ns??? I'm not familiar with that figure of merit.

BTW, my 3080 is running 465.31. Still waiting to catch a WU after the PPA, reboot & reset.

Aren’t you big into folding? ns/day is a very common metric for measuring computation speed in molecular modeling.

---

It’s not real. I’ve manipulated the coproc_info file to report what I want.

Actual driver in use is 460.84

Ok why i ask was that device name is unknown for my 3080Ti and had some hope that driver you used would fix that.
So i could go coproc file and edit instead.

What driver are you using? The 3080ti won’t be detected until driver 460.84. Anything older will not know what GPU that is.

NVIDIA-SMI 465.27 Driver Version: 465.27 CUDA Version: 11.3

Could not use 460 for 3080Ti so i had to move latest ubuntu provided and it would this version.
boinc-client detect name as
Coprocessors NVIDIA NVIDIA Graphics Device (4095MB) driver: 465.27

I edit coproc_info.xml but it does not change when i update to project and if i restart boinc-client it will wipe even if ai change driverversin inside file.

Maybe i could lock file to root only to prevent boinc to write permission but i better not.

You need driver 460.84 for 3080ti. You can use that one.

You can also use 465.31, but that driver is about a month older, 460.84 will be better unless you absolutely need some feature from the 465 branch.

---

Greger, I just can't get my head around what it means. So out of the 8.64E13 ns in a day you only calculate for 159 ns??? I'm not familiar with that figure of merit.

BTW, my 3080 is running 465.31. Still waiting to catch a WU after the PPA, reboot & reset.

As mention before here it is possible time the device could genereate a folding event for that device but you need take in count the complexity of folding time in and amount of atoms have big affect on it and possible other parameters in modelling event.

Think of see it as a box and you have x y z and it build up protein with atoms then make fold of it. In total result it would be very very short event.

There was a free tool before and possible available still today that you could use to open data that done directly with after it was done. users have done this at folding@home and posted in forums.

Not sure if that is free for acemd

You need driver 460.84 for 3080ti. You can use that one.

You can also use 465.31, but that driver is about a month older, 460.84 will be better unless you absolutely need some feature from the 465 branch.

ok thanks