Message boards : News : GPU computing is opening to everyone
Message board moderation
| Author | Message |
|---|---|
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
Until now, GPU tasks were only sent to computers we enabled one by one. From now on, every compatible computer will receive them automatically, with nothing to request. Who will receive GPU tasks • Computers with an NVIDIA graphics card, on Windows or Linux, with a recent driver (528 or newer on Windows, 525 or newer on Linux). • Each GPU task is sized for your exact card model. • Cards we don't know yet, and computers mixing several different card models, won't receive GPU tasks for now. They keep their usual CPU work. If you don't want GPU tasks Your choice is always respected. Any one of these options works: 1. For all your computers: open your Eternity@home project preferences, click Edit preferences and uncheck "Use NVIDIA GPU". Your CPU work continues normally. 2. For one computer only: in BOINC Manager, menu Activity, choose "Suspend GPU". 3. Only while you use your computer: in BOINC Manager, Options → Computing preferences, uncheck "Use GPU while computer is in use". If you have already turned GPU off, nothing changes for you. Thank you all for your support. Questions are welcome in the forum! |
|
Send message Joined: 15 Oct 26 Posts: 1 Credit: 3,924,703 RAC: 36,678 |
It would have been nice for you to set the default to not using the GPU. I had other things going on on this system that were interupted/delayed. Cheers Waffle |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to waffleironhead's message of 27 Oct 2026: It would have been nice for you to set the default to not using the GPU. I had other things going on on this system that were interupted/delayed. Sorry for the disruption, Waffle, and thanks for telling us. BOINC uses every resource a project supports unless the preference says otherwise, which is why GPU was on by default. In hindsight, we should have given more notice before switching it on. To turn it off, uncheck "Use NVIDIA GPU" in your project preferences. Your CPU work continues normally, and any GPU task already downloaded can simply be aborted. |
dskagcommunitySend message Joined: 28 Sep 26 Posts: 31 Credit: 8,406,604 RAC: 229,439 |
Do i have to reset the project? i still get privates sent after beta so i think its on serverside or so? And i dont know if they search for real boards. https://boinc.eternityathome.org/eternity/workunit.php?wuid=3758785 as example ---------- 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 27 Oct 2026: Do i have to reset the project? i still get privates sent after beta so i think its on serverside or so? And i dont know if they search for real boards. No reset needed, your machine is fine. The "private" in the name is just a leftover from the beta. During the beta, each enabled computer had its own GPU version, and a few tasks were still created under that name. They are real search tasks: the same campaign as all the other GPU tasks, validated with full credit. Your WU 3758785 validated normally. Your RTX 5070 Ti is eligible for the regular public GPU tasks, so there's nothing to change on your side: keep "Use NVIDIA GPU" ticked and new tasks allowed. If you get no GPU work at all over the next few hours, let me know and I'll look at it. |
|
Send message Joined: 28 Sep 26 Posts: 17 Credit: 15,948,746 RAC: 194,165 |
Just started crunching gpu production tasks. I've two questions based on my limited observation. 1. For AVX-512 capable cpu, will there be a speed improvement if the cpu processing is done specifically with AVX-512 as opposed to AVX2 instructions? 2. It seems that a maximum of 4 threads are been used. Can the number of threads be selected by users? Maybe an option to select this is provided in the user account project preferences? If adding more threads will help to speed up the gpu run time, might be worth looking at this. I've one comment: If AVX2 instruction are mostly used throughout the gpu processing, best to reserve full physical cores rather than cpu logical threads to speed up the gpu processing. Is that an accurate statement or perhaps the AVX2 instruction used is only a small fraction of the total processing time? |
dskagcommunitySend message Joined: 28 Sep 26 Posts: 31 Credit: 8,406,604 RAC: 229,439 |
Ohhhh okai, good to know. thx ^^ When it can be easy done, you could remove them for discspace cos one of my boinccluster is offline for about 2 weeks, in this one was some of the testing maschines too. ---------- 24/7 Crunching since 2011 ----------- DSKAG Austria: http://www.dskag.at
|
|
Send message Joined: 28 Sep 26 Posts: 14 Credit: 56,385,638 RAC: 2,119,212 |
Hi! Seems to be out of GPU work! |
Henk HaneveldSend message Joined: 28 Sep 26 Posts: 1 Credit: 326,396 RAC: 6,587 |
Why is this application set to use 2 CPU's next to the GPU. It only used 1 CPU, the other was idle. Eternity II NVIDIA GPU 273 (4096 tickets) v2.73 (cuda_e2_pubv_s2_legacy_win_4096_60d9f02e) windows_x86_64 I have created an app_config.xml file for the project to only use 1 CPU. The current WU running seems to do fine with just that. |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
@Hans Sveen: you're right, and sorry about that. Between about 01:10 and 09:30 UTC this morning, the server stopped creating new GPU tasks because of a stuck safety check. It's fixed, and we've added an automatic recovery for that case. New GPU tasks are going out again, and your RTX 3080 Ti already received two. @Henk Haneveld: good observation. The 2 CPUs are a reservation sized for fast cards: for most of the task one core feeds the GPU, but at the end of each wave the task finishes the last work on extra CPU threads. On a fast card that goes above 2 cores at times; on a smaller card like yours one core is mostly enough. Your app_config.xml with 1 CPU is perfectly fine: the task stays correct, only that short end phase may take a bit longer. @pututu: thanks for the careful look! • AVX-512: no gain expected. The search is branchy integer work (placing pieces and checking edges), not vector math. We tested an AVX2 build and it was actually slightly slower than the generic one. • Threads: the up to 4 threads you see are the end-of-wave phase described above. Most of the work runs on the GPU, so more threads would only shorten that short final phase. It isn't user-selectable at the moment, but I've noted the idea. • Physical cores vs logical threads: since the code doesn't lean on AVX, logical threads work fine and there's no need to reserve full physical cores. @dskagcommunity: thanks for the suggestion, noted. Nothing to do on your side: tasks sent to computers that stay offline simply expire at their deadline and are sent to someone else, so nothing is lost. |
|
Send message Joined: 29 Sep 26 Posts: 4 Credit: 2,503,434 RAC: 77,792 |
Can we expect WU for AMD GPU ? |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to franky82's message of 28 Oct 2026: Can we expect WU for AMD GPU ? Thanks for offering your RX Vega 64, franky82 (this also answers your post in the beta thread)! Good news: an OpenCL version of the GPU app, which will run on AMD cards, is planned. I don't have a release date yet, and it will need its own testing before it goes live, so we'll announce it in the News forum when it's ready. For now the GPU app runs only on NVIDIA cards, but in the meantime that computer's CPU can crunch the regular CPU tasks, which help just as much. |
|
Send message Joined: 28 Sep 26 Posts: 17 Credit: 15,948,746 RAC: 194,165 |
@benjamin:Thanks for the clarification about the avx2.I saw the program engine name "bw_nvidia_avx2_0.4" and thought that avx2 is widely used. Perhaps you forgot to change the name since you said that when tested on avx2 build it was slightly slower, so I assume a generic one was used currently to improve the speed slightly? On the number of threads used, I notice that four threads are also used even if it is not at the end-of-wave phase but during the mid-phase of the gpu processing. I monitor this with the linux top command and the %CPU header is showing ~400% cpu activity instead of 100% if one thread is used. The TIME+ field is the total accumulated cpu running time During the first 10 mins of running based on elapsed time shown in the boinc manager, the %CPU is running at 100% There was a spike of more than 100% CPU activity during the initial startup. PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 4449 root 39 19 16.5g 243820 143016 S 100.0 0.7 11:42.86 e2_bw_nvidia_av At around 17 mins of elapsed time, %CPU is running close to 400% (i.e. 4 threads) PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 4449 root 39 19 16.6g 276640 143016 S 381.8 0.8 26:52.15 e2_bw_nvidia_av I'm now logging the output of the linux top command every 30 seconds starting from 33 mins gpu run time (which is about halfway of a typical runtime of the gpu task ~1.0- 1.5 hours on my 5070). Here is the task that I monitored https://boinc.eternityathome.org/eternity/result.php?resultid=4095368. Somehow for this task, the end phase is using 1 thread but I've seen 4 threads on the other two tasks that I successfully completed. Note: the stderr shows total task run time = 3,677.0 and total cpu run time = 10,670.26, indicating an average of 2.90 threads were been used to complete this task. 4449 root 39 19 17.0g 745380 143016 S 410.0 2.3 96:11.95 e2_bw_nvidia_av 4449 root 39 19 17.0g 766336 143016 S 408.7 2.3 98:14.69 e2_bw_nvidia_av 4449 root 39 19 17.1g 784420 143016 S 408.6 2.4 100:17.40 e2_bw_nvidia_av 4449 root 39 19 17.1g 801496 143016 S 408.4 2.4 102:20.09 e2_bw_nvidia_av 4449 root 39 19 17.1g 823912 143016 S 409.1 2.5 104:22.95 e2_bw_nvidia_av 4449 root 39 19 17.1g 845212 143016 S 409.3 2.6 106:25.85 e2_bw_nvidia_av 4449 root 39 19 17.1g 865524 143016 S 409.2 2.6 108:28.74 e2_bw_nvidia_av 4449 root 39 19 17.1g 887492 143016 S 409.2 2.7 110:31.61 e2_bw_nvidia_av 4449 root 39 19 17.1g 909640 143016 S 409.2 2.8 112:34.48 e2_bw_nvidia_av 4449 root 39 19 17.2g 930228 143016 S 408.4 2.8 114:37.12 e2_bw_nvidia_av 4449 root 39 19 17.2g 948872 143016 S 408.8 2.9 116:39.88 e2_bw_nvidia_av 4449 root 39 19 17.2g 964012 143016 S 410.1 2.9 118:43.04 e2_bw_nvidia_av 4449 root 39 19 17.3g 983668 143016 S 410.8 3.0 120:46.40 e2_bw_nvidia_av 4449 root 39 19 17.3g 982.7m 143016 S 412.2 3.1 122:50.18 e2_bw_nvidia_av 4449 root 39 19 17.3g 1.0g 143016 S 413.4 3.1 124:54.32 e2_bw_nvidia_av 4449 root 39 19 17.3g 1.0g 143016 S 413.7 3.2 126:58.54 e2_bw_nvidia_av 4449 root 39 19 17.3g 1.0g 143136 S 412.3 3.3 129:02.35 e2_bw_nvidia_av 4449 root 39 19 17.3g 1.0g 143136 S 409.7 3.3 131:05.37 e2_bw_nvidia_av 4449 root 39 19 17.3g 1.1g 143136 S 409.9 3.4 133:08.45 e2_bw_nvidia_av 4449 root 39 19 17.4g 1.1g 143136 S 409.6 3.5 135:11.45 e2_bw_nvidia_av 4449 root 39 19 17.4g 1.1g 143136 S 409.2 3.5 137:14.30 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.1g 143136 S 409.7 3.6 139:17.33 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 409.5 3.7 141:20.29 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 409.1 3.8 143:23.14 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 409.1 3.8 145:25.98 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 409.2 3.9 147:28.85 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 409.2 4.0 149:31.74 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 408.6 4.0 151:34.43 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 408.7 4.0 153:37.17 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 408.7 4.0 155:39.89 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 408.3 4.0 157:42.51 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 408.0 4.0 159:45.03 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 407.3 4.0 161:47.33 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 406.6 4.0 163:49.42 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 405.8 4.0 165:51.29 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 215.0 4.0 166:55.85 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 106.2 4.0 167:27.73 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 104.7 4.0 167:59.16 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 107.8 4.0 168:31.53 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 108.0 4.0 169:03.95 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 105.4 4.0 169:35.58 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 104.9 4.0 170:07.05 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 104.6 4.0 170:38.45 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 103.2 4.0 171:09.43 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 103.4 4.0 171:40.44 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 103.6 4.0 172:11.52 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 103.2 4.0 172:42.49 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 103.1 4.0 173:13.44 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 104.1 4.0 173:44.70 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 103.8 4.0 174:15.84 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 102.7 4.0 174:46.65 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 102.8 4.0 175:17.49 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 102.2 4.0 175:48.17 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 102.8 4.0 176:19.05 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 101.9 4.0 176:49.64 e2_bw_nvidia_av 4449 root 39 19 17.5g 1.2g 143136 S 102.0 4.0 177:20.27 e2_bw_nvidia_av |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
@pututu: thank you for the careful measurements: you're right on both points, and I need to correct my earlier answer. • AVX2: the CPU part of the GPU app does use AVX2. The "avx2" engine is selected automatically when the CPU supports it, and on difficult work it is about 3 times faster than the generic build. The "slightly slower" result I mentioned came from an older CPU-only engine, not from this app. Sorry for the confusion. • AVX-512: we haven't built or tested an AVX-512 version yet, so I can't say whether it would help. It's a fair idea to try later. • 4 threads: also correct. The GPU app runs 4 CPU threads that solve the hardest endgames handed over by the GPU, and they work during the whole task, not only at the end of a wave. Your top log matches exactly: about 400% while the GPU is busy, then about 100% for the final phase. Your 2.9 average threads (10,670 s CPU for 3,677 s runtime) is typical for a fast card. The thread count isn't user-selectable at the moment. • Physical cores: since AVX2 is used, sharing a physical core with another heavy task could slow these threads down a bit. We haven't measured it, though. @Henk Haneveld: a small correction to what I told you too. Your observation is right for a smaller card: it hands fewer endgames to the CPU, so about 1 core is used. Your app_config.xml with 1 CPU remains fine. |
|
Send message Joined: 28 Sep 26 Posts: 17 Credit: 15,948,746 RAC: 194,165 |
benjamin: Thanks for the update. Between AVX-512 vs number of threads, I would probably focus on the latter as I'm noticing that faster card like 4090, 5080 and 5090 are not finishing the tasks any faster than my 5070. It could be that the limitation is on the CPU side, either not allocating full physical cores or perhaps not having enough cores to finish the CPU workload quickly and feed the GPU. Just my guess. In the hybrid approach for solving this puzzle with GPU, having a balance between CPU and GPU appears to be quite important. Let the users have the option to play with the number of threads on their system rather than fixing it to 4 threads max. |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to pututu's message of 28 Oct 2026: benjamin: Thanks for the update. Between AVX-512 vs number of threads, I would probably focus on the latter as I'm noticing that faster card like 4090, 5080 and 5090 are not finishing the tasks any faster than my 5070. It could be that the limitation is on the CPU side, either not allocating full physical cores or perhaps not having enough cores to finish the CPU workload quickly and feed the GPU. Just my guess. In the hybrid approach for solving this puzzle with GPU, having a balance between CPU and GPU appears to be quite important. Let the users have the option to play with the number of threads on their system rather than fixing it to 4 threads max. @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. Part of that is noise: some hosts run two GPU tasks per card at once, some CPUs are busy with other projects, and the number of hard endgames varies a lot from task to task. But the top cards clearly need a better fit. The plan is bigger work units for the fastest cards and a new version of the GPU app tuned for them, and your suggestion about the thread count is noted for that version too. Right now our priority is the work distribution and, above all, saving and resuming GPU tasks, so that a reboot or a pause doesn't lose progress. The big-card version comes after that. I can't give a date yet, and we'll announce it in the News forum. |
|
Send message Joined: 27 Oct 26 Posts: 5 Credit: 91,857 RAC: 6,312 |
A little off-topic. Beyond the fact that this whole project is interesting and exciting, what really makes it appealing is that Benjamin goes into such detail. Most BOINC projects lack this level of communication from the administrators, so even if a project seems useful or interesting to me, I lose interest when the admin is unreachable and doesn’t provide any information or show any signs of life. As for me, I’m using a very old PC (i5-4460 + GTX750Ti), but I’m still hopeful for small successes, and I’m eagerly awaiting the first 463 boards. It’s a bit like buying a lottery ticket with every single work unit. :) Keep up the great work, Benjamin! |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to C0d3r's message of 29 Oct 2026: A little off-topic. Thank you so much, C0d3r, that really means a lot! Answering questions here is part of the fun for me, and many of the improvements of the last weeks came straight from your reports. Good news about your old PC: your GTX 750 Ti has already completed 2 GPU tasks successfully, and a third one is running. Old hardware is perfectly welcome. And you're right, every task is a lottery ticket: each one has the same chance to hold a 463, whatever machine computes it. Fingers crossed for your first one! |
|
Send message Joined: 27 Sep 26 Posts: 42 Credit: 141,621,123 RAC: 4,172,786 |
While I'm happy it's open to more users, I wish all apps are available to everyone. Or at least GPUs with larger amounts of memory can run all the apps including the apps with lower memory requirements. |
|
Send message Joined: 23 Sep 26 Posts: 242 Credit: 54,980,907 RAC: 1,491,573 |
In reply to mmonnin's message of 29 Oct 2026: While I'm happy it's open to more users, I wish all apps are available to everyone. Or at least GPUs with larger amounts of memory can run all the apps including the apps with lower memory requirements. Thanks mmonnin, I understand the wish. The GPU tasks aren't split by memory but by speed: each card gets a batch size chosen so that a task takes roughly the same time on every model. A fast card on a small batch would spend more of its time starting and finishing, and less time searching. Opening more apps also wouldn't give your hosts more GPU work, since the limit of 2 GPU tasks per card counts all GPU apps together. Right now all your GPU hosts hold their 2 tasks per card. The one real benefit would be a fallback if the batch size for your card ever runs out, and I've noted that idea. |
