Username: SkyHighWeFly
1 host , Host ID 2003
RTX 4070 Super; 12281 MB RAM; Driver: 61714; Windows 11

@Bill F: thanks for the report, great to see the GTX 1660 SUPER validating!

@Landjunge and @Hans Sveen: sorry, this was on our side. A GPU task got stuck on each of your hosts, and after it was aborted, the server put the host on hold as a safety measure, so it stopped receiving GPU work. Your machines did nothing wrong. The block has now been cleared for both:
• Landjunge: BOINC has already sent two new GPU tasks to host 353.
• Hans: a new 65,552-ticket GPU task is ready for your 4080 (host 166). Please click Update in BOINC to fetch it.
The tickets from the aborted tasks are kept and will be recomputed separately; nothing is lost.

@Werinbert: fair question. Each GPU task reserves 6 CPU threads because the deepest endgames are computed on the CPU and arrive in bursts. Most of the time only about one thread is busy, as Dargo75 and Freewill also saw. We're looking at the measurements from this beta to see how much really needs to be reserved.

@SkyHighWeFly: thanks for volunteering! Your RTX 4070 SUPER (host 2003) is now enabled, and your first 65,552-ticket GPU task is ready. Please click Update in BOINC to fetch it.

Username: kotenok2000
1 host: 127
NVIDIA GeForce GTX 1650; 4095MB RAM; Driver: 61692; OpenCL: 300 Windows 11

@Landjunge and @Hans Sveen: sorry, this was on our side. A GPU task got stuck on each of your hosts, and after it was aborted, the server put the host on hold as a safety measure, so it stopped receiving GPU work. Your machines did nothing wrong. The block has now been cleared for both:
• Landjunge: BOINC has already sent two new GPU tasks to host 353.
• Hans: a new 65,552-ticket GPU task is ready for your 4080 (host 166). Please click Update in BOINC to fetch it.
The tickets from the aborted tasks are kept and will be recomputed separately; nothing is lost.


Yes, got this one https://boinc.eternityathome.org/eternity/workunit.php?wuid=3418052 late afternoon,
but since then no new wu!
I have tried/trying update but still no new wu!

Hans S.

I have a task (on host 30) that seems to be stuck. The GPU and CPU are both active but the task sits at 99.0419819380% done and completed 64924 of 65552. Should I abort it? I am afraid of doing so as it may stop me from getting new work per earlier messages in this thread.

Ive now added the project to BOINC on that machine, it should communicate and be visible soon.

The host name is Test

Thanks

Hi!
If useful I also noticed that the last wu on my 4080 card used app.version 2.68 and the 3080
has been running app. version 2.72.
This change appeared after 00.00 UTC.

The Boinc Manager's Event log also did notice this:

22.10.2026 02.34.34 | Eternity II | Sending scheduler request: To send trickle-up message.
22.10.2026 02.34.34 | Eternity II | Requesting new tasks for NVIDIA GPU
22.10.2026 02.34.36 | Eternity II | Scheduler request completed: got 0 new tasks
22.10.2026 02.34.36 | Eternity II | Project requested delay of 61 seconds
22.10.2026 02.35.54 | | Project communication failed: attempting access to reference site
22.10.2026 02.35.55 | | Internet access OK - project servers may be temporarily down.
22.10.2026 02.39.42 | Eternity II | Sending scheduler request: To send trickle-up message.
22.10.2026 02.39.42 | Eternity II | Requesting new tasks for NVIDIA GPU
22.10.2026 02.39.43 | Eternity II | Scheduler request completed: got 0 new tasks
22.10.2026 02.39.43 | Eternity II | Project requested delay of 61 seconds
22.10.2026 02.47.39 | Eternity II | Sending scheduler request: To send trickle-up message.


Hans S

In reply to benjamin's message of 21 Oct 2026:

In reply to mmonnin's message of 21 Oct 2026:

Then GPUs are not worth running the small tasks. Power wise a CPU will be better.

Also why do the beta CPU tasks have a 2 hour deadline? It interrupts anything else that is running

You're right that the small ones are poor value, and FritzB is right about the per-ticket rate (thanks FritzB!). The small tasks you got were test and recovery tasks: a few hundred tickets, plus several minutes of fixed start-up, so per hour they pay badly. They are not the normal size. Your two recent full-size tasks (65,650 tickets) each earned 87,825 credits in about 3 h 10 on the 3070 Ti, roughly 27,600 credits per hour. For comparison, the same 32-thread machine on CPU work would earn around 9,600 per hour on our CPU basis.

About the short deadline: the CPU tasks that finish the last tickets of a GPU batch are kept short on purpose, so a batch doesn't stay open for days. You're right that it pushes other work aside; new ones now have a 4-hour deadline instead of 2.

Please add a Beta option in the user preferences where users must opt in. Unless the queue is like 0.1 days the beta CPU tasks interrupt other tasks.

Why is it that one of my computers (ID 31, RTX 4060ti) keeps getting 32k ticket tasks and my other computer (ID 30, RTX 5060) keeps getting 64k ticket tasks?

In reply to mmonnin's message of 22 Oct 2026:

In reply to benjamin's message of 21 Oct 2026:

In reply to mmonnin's message of 21 Oct 2026:

Then GPUs are not worth running the small tasks. Power wise a CPU will be better.

Also why do the beta CPU tasks have a 2 hour deadline? It interrupts anything else that is running

You're right that the small ones are poor value, and FritzB is right about the per-ticket rate (thanks FritzB!). The small tasks you got were test and recovery tasks: a few hundred tickets, plus several minutes of fixed start-up, so per hour they pay badly. They are not the normal size. Your two recent full-size tasks (65,650 tickets) each earned 87,825 credits in about 3 h 10 on the 3070 Ti, roughly 27,600 credits per hour. For comparison, the same 32-thread machine on CPU work would earn around 9,600 per hour on our CPU basis.

About the short deadline: the CPU tasks that finish the last tickets of a GPU batch are kept short on purpose, so a batch doesn't stay open for days. You're right that it pushes other work aside; new ones now have a 4-hour deadline instead of 2.

Please add a Beta option in the user preferences where users must opt in. Unless the queue is like 0.1 days the beta CPU tasks interrupt other tasks.

Please remove my CPU hosts from Beta, Hosts 8 to 15. So much time is being wasted by the task interruptions and WUProp hours lost as it's not recognizing the different version of the app.

Update for all GPU testers: all GPU hosts now run our newest versions (2.72 for modern cards, 2.73 for older ones), with two improvements: an anti-stall safeguard (if the GPU stays idle for 90 seconds near the end of a task, the remaining tickets go to CPU tasks right away), and an end-of-task handoff at 1% of tickets instead of 0.5%, which makes overall GPU throughput about 7.5% faster. We also restarted the server this morning; GPU work supply was slow for a while afterwards, sorry about that.

@Werinbert: we've prepared your computer to accept the cancellation of that one old 2.68 task without blocking further GPU work. Please abort only that task (workunit 3418165), leave your other tasks running, keep new tasks allowed and click Update. Its unreturned work stays reserved for a later recovery, but this aborted task won't earn credit, sorry about that. Your computer already completed tasks with the new version, which handles the end of tasks properly. About task sizes: each card class gets a batch size matched to its speed. Your RTX 5060 gets 65,536 tickets per task, while the RTX 4060 Ti currently gets 32,768 under our class rules. These sizes will be refined as we measure more cards.

@Hans Sveen and @SkyHighWeFly: your hosts are enabled and set up correctly on our side, but they haven't been sent new GPU work yet. This is on our side, not your settings, and we're looking into it.

@mmonnin: good suggestion, thanks. An opt-in beta setting for these short CPU end-of-task jobs makes sense, and we've passed it on to the team.

@BorgProcessor: thanks, we can see your V100 machine now! It still reports driver 528.89, though, which is too old for our application. Once you've updated to a recent NVIDIA driver (560 or newer), let us know and we'll enable it.

@kotenok2000: thanks for volunteering! Your GTX 1650 (host 127) is now enabled for the GPU beta. Just keep new tasks allowed and click Update in BOINC; no CUDA installation or reset is needed.

Updated to driver version 560.94, cheers.

In reply to mmonnin's message of 22 Oct 2026:

In reply to mmonnin's message of 22 Oct 2026:

In reply to benjamin's message of 21 Oct 2026:

In reply to mmonnin's message of 21 Oct 2026:

Then GPUs are not worth running the small tasks. Power wise a CPU will be better.

Also why do the beta CPU tasks have a 2 hour deadline? It interrupts anything else that is running

You're right that the small ones are poor value, and FritzB is right about the per-ticket rate (thanks FritzB!). The small tasks you got were test and recovery tasks: a few hundred tickets, plus several minutes of fixed start-up, so per hour they pay badly. They are not the normal size. Your two recent full-size tasks (65,650 tickets) each earned 87,825 credits in about 3 h 10 on the 3070 Ti, roughly 27,600 credits per hour. For comparison, the same 32-thread machine on CPU work would earn around 9,600 per hour on our CPU basis.

About the short deadline: the CPU tasks that finish the last tickets of a GPU batch are kept short on purpose, so a batch doesn't stay open for days. You're right that it pushes other work aside; new ones now have a 4-hour deadline instead of 2.

Please add a Beta option in the user preferences where users must opt in. Unless the queue is like 0.1 days the beta CPU tasks interrupt other tasks.

Please remove my CPU hosts from Beta, Hosts 8 to 15. So much time is being wasted by the task interruptions and WUProp hours lost as it's not recognizing the different version of the app.

Thanks mmonnin, you were right, and your feedback helped us fix it. We checked the numbers: the short CPU end-of-task jobs paid about 2.5 times less per CPU hour than a normal CPU task, because they contain the hardest tickets of each GPU batch. We've now adjusted their credit so that an hour of these short tasks earns the same as an hour of normal CPU work: the rate goes up about 2.5 times, from 1.55 to 3.85 credits per ticket. This applies to everyone, for the end-of-task jobs coming from GPU batches created from now on; jobs from older batches keep their previous rate. A new GPU batch and its CPU end-of-task jobs still need to finish first, so I expect the first increased credits in about 2 to 6 hours, depending on the work already downloaded and on how long the calculations take; that's an estimate, not a guaranteed time.

If you don't mind, we'll leave your hosts as they are for now, since I think the real problem was an obvious lack of credit. Let me know if the interruptions are still a problem once the new rate kicks in.

cgrinch

At last, whatever You did my 4080 is now happy crunching a new unit (App.version 2.72)!

Thank You 😉👍

Hans S.

@BorgProcessor: thanks for updating the driver! GPU beta access is now enabled for your V100 host (#2005). Please keep new tasks allowed and click Update in BOINC Manager; version 2.72 will download automatically. Since that PC has 4 CPU threads, the GPU task reserves 3 of them and leaves one free. Let us know how the first GPU task runs!

@cgrinch: thanks for joining! We can see two NVIDIA hosts on your account (an RTX 5060 Laptop and an RTX 3060 Ti), but neither has contacted the project since 12 October. Please start BOINC on the one(s) you'd like to use for the beta and let us know, and we'll enable them.

@Hans Sveen: great news, thanks for letting us know, and thanks for your patience!

Well I let the stalled task run over night, it apparently actually is doing something but at a very slow speed. Over the last 6 hours it went from 99.324 to 99.385% done. Total time spent on the task is over12 hours. Going to abort as this will take for ever to get done.

I'm seeing a lot of variation in the runtime for v2.72 task for the exact same credit. Spans 6,320 to 12,450 seconds for 43,912.73 credits. Is that normal or a possible issue?

https://boinc.eternityathome.org/eternity/results.php?userid=117&offset=0&show_names=0&state=0&appid=

@Werinbert: thanks, and sorry it dragged on that long. Aborting it was the right call: your host kept its GPU access and received a new task right away, on the new version that no longer stalls like this. The tickets from the aborted task are kept for a later recomputation.

@Freewill: yes, that's normal. Every task of that size has the same number of tickets, so it earns the same credit, but some batches contain many more deep endgames than others, and those take much longer to compute. Sharing the card with other work, for example through MPS, also changes the runtime. Over many tasks it evens out. If one task runs several times longer than your usual ones, though, let us know.

Thanks, benjamin. Just to update you, I have not been running MPS for the last 1.5 days and I never shared the card with other tasks, even with MPS, since Eternity II tasks request 1.0 GPU by default.