I take the Question back.

Could I get these enabled,

LCB001

ID 2037
3060
616.64
Win 10

ID 2038
1660 Super
610.88
Win 11

Thanks

My tasks (Eternity II NVIDIA GPU 272 (32768v tickets)) on Computer 30 were completing to 99% and then quickly handing the task back to the server. Now the tasks are completing to 99% and then sit a around for a while before returning the task to the server. Why is this happening?

@LCB001: thanks for joining! Both hosts are now enabled for the GPU beta: the RTX 3060 (2037) and the GTX 1660 SUPER (2038). Just keep new tasks allowed and click Update in BOINC; no CUDA installation or reset is needed.

@Werinbert: good eye, that's expected. We're running a short test where some GPU tasks keep going a little longer, until about 99.5% instead of 99%, before handing their last tickets to CPU tasks. One of your tasks on computer 30 was part of it. The goal is to send about half as many small CPU end-of-task jobs to other volunteers; so far it costs the GPU little to no extra time. Thanks for noticing and reporting it!

Well in some cases it nearly doubled the processing time (looking at the stderr output) without the corresponding credit increase.

In reply to Werinbert's message of 26 Oct 2026:

Well in some cases it nearly doubled the processing time (looking at the stderr output) without the corresponding credit increase.

Thanks Werinbert, we looked at the logs of your last 14 GPU tasks on computer 30. Good news: the actual run time didn't change. Every task, including the test one, ran 53 to 55 minutes from start to finish. What varies is the "CPU time" at the end of the log, from about 1 to 2.5 hours. That depends on how many deep endgames a batch contains, because those are computed on the CPU threads alongside the GPU, and some batches have many more than others. Since credit is paid per ticket and the run time stays the same, your credit per hour isn't affected.

benjamin, now that the GPU option is being opened up to eveyrone, is it okay to try running 2+ tasks per GPU with MPS to see if there's a benefit? If so, do you (or anyone) have an app_config.xml example for this project?

In reply to Freewill's message of 26 Oct 2026:

benjamin, now that the GPU option is being opened up to eveyrone, is it okay to try running 2+ tasks per GPU with MPS to see if there's a benefit? If so, do you (or anyone) have an app_config.xml example for this project?

Thanks for asking first, Freewill! For now, please keep one GPU task per card. The only time someone tried two tasks at once with MPS (an RTX 4080), one of them crashed near the end, and the app isn't validated for that yet. BOINC already keeps a second task ready for each card, so the GPU never waits between tasks and no app_config.xml is needed. We'll let everyone know if running several tasks per GPU becomes supported.

I run (on a other computer) an AMD Radeon RX Vega 64 if it can help !

I asked the same.
Answer: in the moment just NVIDIA until stable, AMD later.

@franky82: thanks for offering your RX Vega 64! As I wrote in the other thread, an OpenCL version of the GPU app, which will run on AMD cards, is planned, and we're already working on it with rebirther. No release date yet: it will need its own testing first, and we'll announce it in the News forum. Until then, that computer's CPU can crunch the regular CPU tasks.

@vonboedefeldt: exactly right, thanks for passing it on: NVIDIA first while we make it stable, AMD later through OpenCL.

Cool :) When aannnnyyyytime there will be a closed beta again, you can contact me for testing again with old GPUs xDD 2*RX 280 i think they was ^^

---

---------- 24/7 Crunching since 2011 -----------
DSKAG Austria: http://www.dskag.at

In reply to dskagcommunity's message of 29 Oct 2026:

Cool :) When aannnnyyyytime there will be a closed beta again, you can contact me for testing again with old GPUs xDD 2*RX 280 i think they was ^^

Thanks dskag, noted! Your help during the beta was really useful. When the AMD version is ready for testing, we'll announce it in the forum, and we'll gladly take your two RX 280s if the app can run on them. Older cards are a good test, but I can't say yet which AMD generations will be supported.

The GPU task normally run for around an hour, yet I get a fair number of long running multi-hour tasks. The current one has run for over 5 hours and is still crunching. Unfortunately, these long running tasks are granted the same credits as the one hour tasks.

edit: I am getting the following in the stderr output

bw_runner: warning: unable to publish durable progress

In reply to Werinbert's message of 30 Oct 2026:

The GPU task normally run for around an hour, yet I get a fair number of long running multi-hour tasks. The current one has run for over 5 hours and is still crunching. Unfortunately, these long running tasks are granted the same credits as the one hour tasks.

edit: I am getting the following in the stderr output

bw_runner: warning: unable to publish durable progress

Hi Werinbert, thanks for the report, you're right. I checked your Windows host: most GPU tasks take about 1.5 hours, but some take much longer (your longest so far was about 6.5 hours). This is a known issue that we measured just today. At the end of a task, the last few tickets can take a long time to finish, and on Windows this final phase takes about twice as long as on Linux. Your batch size on Windows is the most affected. We haven't found the cause yet, and I won't change the app during the sprint, but it's at the top of my list after it.

About credit: it's fixed per task, so it doesn't reward long tasks. I agree that's unfair to them. Making those long endings shorter is the real fix, and I've noted your point.

The warning unable to publish durable progress is harmless. It only means the file that shows the progress bar couldn't be written at that moment, which can happen briefly on Windows, for example when an antivirus is scanning it. Your results are saved separately and nothing is lost.

I have an NVIDIA RTX PRO 6000 Blackwell Workstation GPU under Win 11 Pro with driver version 32.0.15.8180 (NVIDIA 581.80) DCH / Win11 64 but it won't fetch GPU tasks although I setup BOINC to fetch both CPU & GPU tasks on that machine.

Michael.

---

President of Rechenkraft.net e.V.

In reply to Michael H.W. Weber's message of 30 Oct 2026:

I have an NVIDIA RTX PRO 6000 Blackwell Workstation GPU under Win 11 Pro with driver version 32.0.15.8180 (NVIDIA 581.80) DCH / Win11 64 but it won't fetch GPU tasks although I setup BOINC to fetch both CPU & GPU tasks on that machine.

Michael.

Hi Michael, thanks for reporting it, and what a card! Your BOINC setup was fine. The problem was on our side: each GPU model needs an entry in our list, and the server matches the model name exactly. We had the "RTX PRO 6000 Blackwell Server Edition", but not the "Workstation Edition". It has now been added, so your card should start receiving GPU tasks the next time BOINC asks for GPU work. If nothing arrives within an hour or so, an "Update" of the project in BOINC Manager should help. Otherwise, let me know here. Your AMD card isn't supported for now: an OpenCL version is planned, but I don't have a date yet.

I got a single task on 31 Oct 2026, 3:57:24 UTC but this one shows no progress at all in the BOINC manager since approx. 30 min. now.
Additional estimated runtime ist approx. 1 day 2 hrs.
It utilizes 2 CPU cores plus the GPU but GPU load is virtually zero (temperature around 31°C, task manager indicates 13 of the total 96 GB VRAM as used) with re-occuring peaks (temperature then going up to 79°C, VRAM usage remaining at 13 GB).

So, something is happening but VERY inefficiently: The two CPU cores each show a constant load around 75% on my Threadripper PRO 9955WX (irrespective of the GPU showing load or idleing around).
I think I should cancel the task? Is there any estimation on how long it should really run (initial task runtime estimation in BOINC often is entirely wrong)?

Michael.

Edit: After 50 min of total runtime the task now shows approx. 12.3% progress (~23 hrs left), so I guess I will keep it running - but won't load a second WU until I get feedback from you regarding the inefficient GPU usage and the "hogging" of two CPU cores for that.

2nd edit: After 1 hrs 11 min the progress bar indicates almost 25%, so total runtime could be around 4 hrs 30 min if the tasks behaves linear.

P.S.: I took screenshots of the situations described above but won't post these here. Just send me a PM if these are needed.

---

President of Rechenkraft.net e.V.

In reply to Michael H.W. Weber's message of 31 Oct 2026:

I got a single task on 31 Oct 2026, 3:57:24 UTC but this one shows no progress at all in the BOINC manager since approx. 30 min. now.
Additional estimated runtime ist approx. 1 day 2 hrs.
It utilizes 2 CPU cores plus the GPU but GPU load is virtually zero (temperature around 31°C, task manager indicates 13 of the total 96 GB VRAM as used) with re-occuring peaks (temperature then going up to 79°C, VRAM usage remaining at 13 GB).

So, something is happening but VERY inefficiently: The two CPU cores each show a constant load around 75% on my Threadripper PRO 9955WX (irrespective of the GPU showing load or idleing around).
I think I should cancel the task? Is there any estimation on how long it should really run (initial task runtime estimation in BOINC often is entirely wrong)?

Michael.

Edit: After 50 min of total runtime the task now shows approx. 12.3% progress (~23 hrs left), so I guess I will keep it running - but won't load a second WU until I get feedback from you regarding the inefficient GPU usage and the "hogging" of two CPU cores for that.

2nd edit: After 1 hrs 11 min the progress bar indicates almost 25%, so total runtime could be around 4 hrs 30 min if the tasks behaves linear.

P.S.: I took screenshots of the situations described above but won't post these here. Just send me a PM if these are needed.

Hi Michael, thank you for this very useful report, and you were right to question it. I looked into it: your card is fine. Your first task ran on an older test version of our GPU app, which turns out to be about three times slower than our normal version on the same computers. That version has now been withdrawn.

Your next tasks should use the normal version. For comparison, RTX 5090 cards take about 1 hour per task on it, and you should see much better GPU use. So please do accept a second task: I'd really like to see what your card can do! Some tasks can still take longer because of the slow endings I described above.

The two CPU cores are intended: each GPU task reserves them. The first 10 to 15 minutes of a task also show little progress on every card, so that part is normal.

Your current task is still useful, so I suggest letting it finish, as an aborted task gets no credit. Screenshots aren't needed for now, thanks. If the next task behaves the same way, please let me know here.

The task was valid and took a total of 16,007.29 seconds (i.e. approx 4:45 [hrs:min]) giving 1200 credits (which is not proportional and undercredited compared to CPU tasks runtimes).
Will try a second one.

Michael.