Message boards : News : GPU computing is opening to everyone
Message board moderation
Previous · 1 · 2 · 3 · Next
| Author | Message |
|---|---|
|
Send message Joined: 27 Oct 26 Posts: 5 Credit: 91,857 RAC: 6,312 |
My only issue—though this is probably a problem with my old computer—is that the UI stutters when the GPU is under load, including, for example, watching YouTube videos and scrolling through web pages. I haven’t found a solution yet, and it’s annoying, unfortunately. Here’s what I’ve tried: updating the graphics card driver, disabling GPU acceleration (both in Windows 10 and in the browser)—neither helped. If anyone has any ideas, or if there’s a member here with a similar older setup, I’d appreciate it if they could share their experiences. I’d like to be able to run it even while I’m using the computer. The Task Manager shows 100% utilization—is there a way to control this similarly to how the CPU is managed? I couldn’t find an option for this in the BOINC Manager, only the ability to suspend it when the computer is in use. |
|
Send message Joined: 27 Sep 26 Posts: 42 Credit: 141,621,123 RAC: 4,172,786 |
In reply to benjamin's message of 29 Oct 2026: In reply to mmonnin's message of 29 Oct 2026: Well your apps use like 40% memory as well whether its 8, 12 or 18GB. As I've mentioned before I and others run WUProp and run apps for a certain # of hours. The server/you are choosing which app can be run on our own device. Give the user the option. Allow more than 2. 2x at once works fine btw. The GPU can be kept busy with multiple tasks at once. C0d3r: It's always stuttered Chrome on my W10 desktop running the CPU app like Rakesearch Spectra of Latin squares does. |
|
Send message Joined: 27 Oct 26 Posts: 5 Credit: 91,857 RAC: 6,312 |
The problem is that I'm experiencing lag at the system level, not just in the browser (Brave). I see that there are plans to release an OpenCL version as well. With my old graphics card, the CUDA version ran slower than the OpenCL version on the distributed.net project—I recall the OpenCL version was about 30% faster. I’m still hopeful about this, and I hope I’ll have the chance to test it. |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
@C0d3r: sorry about the stutter. Your GTX 750 Ti is both your display card and a computing card, and our GPU app keeps it fully busy, so the desktop has to wait for it. Unfortunately BOINC can't cap GPU usage the way it caps CPU usage. Two things that may help: • Since your CPU has 4 cores and the GPU app also uses some CPU time, leave one core free: Options → Computing preferences → "Use at most 75% of the CPUs". That often helps with system-wide lag. • If it's still too much, "Suspend GPU computing when computer is in use" lets the card work only while you're away. I've also noted your interest in testing the OpenCL version. It's planned, but I don't have a date yet. @mmonnin: fair points, and thanks for explaining the WUProp angle. Letting volunteers choose their GPU apps and raising the limit of 2 tasks per card are both project settings, and I don't want to change them in the middle of the sprint. I've noted both and will look at them after it. Good to know that running 2 at once works well on your cards. |
|
Send message Joined: 28 Sep 26 Posts: 17 Credit: 15,948,746 RAC: 194,165 |
@benjamin, possible to print the gpu usage to the stderr.txt file? Something similar to einstein@home stderr.txt file here as an example https://einsteinathome.org/task/2035000723. See second last row below "boinc GPU usage set to: 0.330000". It's a good info when trying to compare performances on different cards. Using default boinc setup, this can be read off from the boinc data project directory file: app_config.xml file, otherwise the default is 1 task per gpu. Stderr output <core_client_version>7.23.1</core_client_version> <![CDATA[ <stderr_txt> [10:53:48][1901150][INFO ] Application startup - thank you for supporting Einstein@Home! [10:53:48][1901150][INFO ] Petri's Custom BRP7 v0.41 - CUDA 12.9 - compiled by IanSteveC on Oct 2 2026 at 12:33:43 command_line: ../../projects/einstein.phys.uwm.edu/BRP7_einsteinbinary_x86_64-pc-linux-gnu__cuda1291 -l ../../projects/einstein.phys.uwm.edu/Ter5_3_sband_dns_cfbf00083_segment_5_dms_40.zap -c Ter5_3_sband_dns_cfbf00083_segment_5_dms_40_24.cpt -A 0.04 -P 1.5 -f 500.0 -W --pb_min 3600 --mc_max 1.6 --mp_min 1.1 --mismatch 0.222 --tobs 1800 --alpha 1.0 --start_template_id 8150000 --no_of_templates 50000 -z -d spin -e blocking -i ../../projects/einstein.phys.uwm.edu/Ter5_3_sband_dns_cfbf00083_segment_5_dms_40_24.binary -o ../../projects/einstein.phys.uwm.edu/Ter5_3_sband_dns_cfbf00083_segment_5_dms_40_8000_24_8150000_0_0 [10:53:48][1901150][INFO ] boinc GPU usage set to: 0.330000 [10:53:48][1901150][INFO ] -- |
rebirtherSend message Joined: 24 Sep 26 Posts: 29 Credit: 39,233 RAC: 755 |
The same as for cpu_usage <app_config>
<app>
<name>app name</name>
<gpu_versions>
<gpu_usage>0.33</gpu_usage>
</gpu_versions>
</app>
</app_config>But never tried it. |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
@pututu: good idea, thanks. Writing the GPU usage setting in stderr would make it much easier to compare cards. Changing the GPU app means a new version, and I won't do that during the sprint, but I've added your idea for the next one. @rebirther: thanks for the app_config example, that's the right way to do it. One detail: the project sends at most 2 GPU tasks per card, so even with 0.33 only 2 tasks will run at once. 0.5 gives the same result. |
|
Send message Joined: 28 Sep 26 Posts: 17 Credit: 15,948,746 RAC: 194,165 |
In reply to benjamin's message of 29 Oct 2026: @pututu: good point, and our numbers partly agree with you. Over the last two days, on the largest GPU tasks, the RTX 5080 averages about 96 minutes and the 5070 about 118, but the 4090 about 128, so the fastest cards are not gaining as much as they should. @benjamin: It seems that the Run Time as posted in the Job instance summary section on systems running gpu tasks maybe incorrect.Take one example from your 5070 Ti system: https://boinc.eternityathome.org/eternity/result.php?resultid=5053772. The reported run time for this task is 1 hours 49 min 35 sec as per the Job instance section. However if you look at your stderr output file, the total run time is about 56 mins 15 secs. Boinc wrapper started at 03:29:29 and ended at 04:25:44. Not sure if you use the data from the Job instance section run time to come to above conclusion or whether you did use the run time in the stderr output. It will be good to understand this discrepancy. 2026-11-01 03:29:29 (26612): wrapper: running ../../projects/boinc.eternityathome.org_eternity/bw_runner_nvidia_common_0.17.exe . . . bw_runner: progress 65376/65704 (99%), restored runtime 1695 s 2026-11-01 04:25:44 (26612): bw_runner exited; CPU time 9867.437500 2026-11-01 04:25:44 (26612): called boinc_finish(0) |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to pututu's message of 1 Nov 2026: In reply to benjamin's message of 29 Oct 2026: @pututu: you're right, and good catch. The run time on the task page is wrong for GPU tasks, and my numbers were wrong because of it. Our GPU app uses about two CPU cores, so its CPU time is higher than its real time, but BOINC treats it as a single-thread app. When the server sees less run time than CPU time, it assumes a client bug. It replaces the run time with the CPU time, capped at the time between sending the task and receiving the result. In your example: • real run time: 56 min 15 s, as you found in stderr • CPU time: 9,867 s • time between sent and received: 6,575 s, which is the 1 h 49 min 35 s shown The task also waited about 53 minutes on my PC before it started, and that wait ended up counted as run time. The numbers I gave you came from that field. I've redone them from the stderr times, over the last 48 hours (1,828 GPU tasks). The real average is about 85 minutes, not 118. On the largest tasks: • RTX 5080: about 60 min • RTX 4090: about 61 min • RTX 5070 Ti: about 66 min So your original point holds, even more strongly: the fastest cards gain little over a 5070 Ti. Credit isn't affected, because GPU credit is fixed per task. I can't make changes during the sprint, but it's on my list for after it. |
|
Send message Joined: 28 Sep 26 Posts: 17 Credit: 15,948,746 RAC: 194,165 |
In reply to benjamin's message of 1 Nov 2026: The numbers I gave you came from that field. I've redone them from the stderr times, over the last 48 hours (1,828 GPU tasks). The real average is about 85 minutes, not 118. On the largest tasks: FYI, I'm currently running 5070 with one task per gpu. On the largest tasks (credit around 88K), I'm getting a range of 61 to 67 mins run time based on stderr output file. Yeah, there is something holding back the performance of the high-end cards. There are quite a number of high end cards currently contributing to this search, so I think if you can optimize these cards, more work will get done for the same amount of time. Also, one other suggestion: it may be useful if a user can select the ticket size and the number of cpu cores that they wish to run, at least be able to experiment themselves and figure out a suitable ticket size task to run. Give the users some flexibility and also allow those to use the default values if one just wish to "set and forget". |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to pututu's message of 1 Nov 2026: In reply to benjamin's message of 1 Nov 2026: Thanks, pututu, these are useful numbers. A 5070 at 61 to 67 minutes matches what we see, and it confirms your point: the high-end cards aren't gaining what they should. With so many of them on the project, I agree it's worth improving. Bigger tasks for the biggest cards are one of my priorities, but I don't have a date yet. Your suggestion is noted too: letting volunteers choose the task size and the number of CPU cores, with default values for those who prefer "set and forget". I've added it to my list. I can't promise it yet, as I first need to check what that would change in the way tasks are created and credited. |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
@pututu: thanks for your measurements, they were right and helped a lot. We found why the big cards gained little: the GPU application runs as many tickets at once as the card has compute units × 640 (about 109,000 on an RTX 5090, 82,000 on a 4090, 31,000 on a 5070). A 65,536-ticket task could not fill the biggest cards. So: the RTX 4090/5090 now get bigger tasks (up to 131,072 tickets), the 5070/5060 Ti get 32,768 (shorter tasks, same work per hour), and anyone can tick "also accept the next bigger task size" in the project preferences. For CPU cores, @mmonnin: about your idea of a fallback size if the size for your card runs out: in practice the server keeps a stock of every size and refills it continuously, so it only runs out during an outage, and then all sizes are affected at once. But the new "also accept the next bigger task size" box gives your card a second size anyway, and the 8-tasks-per-card limit is live since 4 Nov. |
dskagcommunitySend message Joined: 28 Sep 26 Posts: 31 Credit: 8,406,604 RAC: 229,439 |
Can someone look into this plz and tell me what was the problem? Its from a 1070ti with 8gb vram as GPU1 on a multiGPU system. I excluded eternity on GPU0 and 2, it instantly errored out and im very sure it should able to run it: https://boinc.eternityathome.org/eternity/result.php?resultid=5911200 Perhaps some multigpu special thingy problem because it is just reported as one of multiple 3060 in this system? its a 32k ticket so it shuld not be a problem with 8GB VRAM on that card. thx! :) ---------- 24/7 Crunching since 2011 ----------- DSKAG Austria: http://www.dskag.at
|
|
Send message Joined: 4 Nov 26 Posts: 6 Credit: 9,506,537 RAC: 442,824 |
hello, "NVIDIA_BETA_INCOMPATIBLE", Could you be having a driver issue? Or perhaps the board isn't supported by the project. |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
@dskagcommunity: I checked the log: the task stopped because of the card generation, not the card itself. BOINC only reports the best card of a machine, so your PC is seen as "RTX 3060" and gets the version built for Ampere cards. That version needs about 56 KB of fast on-chip memory per block, and a Pascal card like the 1070 Ti has 48 KB. We do have a version for the 1070 Ti (the 8,192-ticket tasks), but the server only sends it to machines that report a GTX 10xx card. The way to get both working: run a second BOINC client on that PC that only sees the 1070 Ti. In its cc_config.xml, hide the 3060s with @artik: good reflex to look at the error line. Here it wasn't the driver but the card generation. |
dskagcommunitySend message Joined: 28 Sep 26 Posts: 31 Credit: 8,406,604 RAC: 229,439 |
ok thx for the precise answer! So its like i though. Hm the easiest would be to make a own profile for the host with just this 8k application with the already include of the card for just eternity. Is it planned to make more then 4 profiles? Its always a good idea when a projekt has many Applications. Cos over time the needed possibilities are growing. often have standard as all units, work gpu only, school backup and home CPUs only. And easy, all profile filled with no special ones like now needed ^^ Oh and is there a public table planned witch cards needed for certain ticketsizes? aka like now 1070ti for 8k and 131k for 5070ti etc.? Then you give the oportunity for mixed systems like mine to resolve the problems for themselfs :) ---------- 24/7 Crunching since 2011 ----------- DSKAG Austria: http://www.dskag.at
|
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to dskagcommunity's message of 12 Nov 2026: ok thx for the precise answer! So its like i though. Hm the easiest would be to make a own profile for the host with just this 8k application with the already include of the card for just eternity. @dskagcommunity: yes, a second client that only sees the 1070 Ti is the way to go. You don't even need a special profile for it: since that client reports a GTX 1070 Ti, the server only sends it the 8,192-ticket tasks. A profile can only filter which applications are allowed; it can't make the server see a card that BOINC doesn't report. About profiles: the 4 locations (default, home, school, work) are fixed by BOINC itself. A project can't add more, sorry. A public table of cards and task sizes is a good idea. The home page lists the families for now. For 8,192 tickets on NVIDIA: GTX 1070, 1070 Ti, 1080, 1080 Ti, 980 Ti, TITAN Xp, GTX 1650 / 1660 (not the 1660 Ti), RTX 3050 and Tesla T4. |
dskagcommunitySend message Joined: 28 Sep 26 Posts: 31 Credit: 8,406,604 RAC: 229,439 |
In reply to benjamin's message of 12 Nov 2026: @dskagcommunity: yes, a second client that only sees the 1070 Ti is the way to go. You don't even need a special profile for it: since that client reports a GTX 1070 Ti, the server only sends it the 8,192-ticket tasks. A profile can only filter which applications are allowed; it can't make the server see a card that BOINC doesn't report. About profiles: the 4 locations (default, home, school, work) are fixed by BOINC itself. A project can't add more, sorry. A public table of cards and task sizes is a good idea. The home page lists the families for now. For 8,192 tickets on NVIDIA: GTX 1070, 1070 Ti, 1080, 1080 Ti, 980 Ti, TITAN Xp, GTX 1650 / 1660 (not the 1660 Ti), RTX 3050 and Tesla T4.[/quote] Oh i though it would be enough to disable all others than the 8k Application, thats how it normaly works on other projects. And the 4 locations are expandable, there are several projects already doing it. ---------- 24/7 Crunching since 2011 ----------- DSKAG Austria: http://www.dskag.at
|
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to dskagcommunity's message of 12 Nov 2026: In reply to benjamin's message of 12 Nov 2026: Oh i though it would be enough to disable all others than the 8k Application, thats how it normaly works on other projects. And the 4 locations are expandable, there are several projects already doing it.[/quote] @dskagcommunity: on most projects one application runs on any card, so unchecking the others is enough. Here each task size is built for a given card generation, and the server picks it from the card model BOINC reports. BOINC reports one model per brand for the whole machine, so yours is seen as an "RTX 3060". With only the 8k application checked, the server finds no version for an RTX 3060 and sends nothing. That's why the second client is needed. About locations: as far as I know the BOINC client only knows these 4, but I may be wrong. If you remember which project offers more, tell me and I'll look at how they do it. |
|
Send message Joined: 27 Sep 26 Posts: 6 Credit: 118,076 RAC: 2,573 |
About locations: as far as I know the BOINC client only knows these 4, but I may be wrong. If you remember which project offers more, tell me and I'll look at how they do it. I can think of two projects off the top of my head. PrimeGrid has the following locations / "venues": Default, Home, School, Work, Sun, Mercury, Venus, Earth, Mars, Jupiter, Saturn, Uranus, Neptune, Pluto. BitBoinc has the following locations/ "venues": Default, Home , School, Work, Apollo, Athena, Hermes, Artemis, Zeus, Hera, Poseidon, Demeter, Aristotle, Plato, Socrates, Archimedes, Pythagoras, Hypatia, Euclid, Hippocrates, Ares, Aphrodite, Hephaestus, Dionysus, Hestia, Homer, Herodotus, Thales, Democritus, Eratosthenes |
