Message boards :
Number crunching :
CUDA Versions
Message board moderation
Author | Message |
---|---|
Matt Send message Joined: 6 Jun 10 Posts: 18 Credit: 1,434,540 RAC: 0 |
Hello Everyone, Just curious, what determines the CUDA version WU one's computer receives? I have a rig with two GTX 780Ti cards which I've only ever seen receive CUDA42 WUs and I also have a PC with two GTX 680 cards which I've seen receive CUDA32, CUDA42, and CUDA50. I always keep both machines up to date with the latest WHQL drivers. Both run Win7 x64. Thank you to anyone who can take the time to help. |
Darth Beaver Send message Joined: 20 Aug 99 Posts: 6728 Credit: 21,443,075 RAC: 3 |
Hello Matt Cuda versions are based on the type of card you have but also if you run Lunatics you can speifie what cuda version you wish to use . Your cards are both capable of running cuda 50 It sounds like your using stock Bionic so others will have to explain why it's using version 32 or version 42 or version 50 Basiclly any card under i think a gt200 you use cuda 32 any card under gtx 490 use cuda 42 and gtx 500 or above or with a Keplar version chip use cuda 50 but others will help you a bit more with info so just wait and when they got time i'm shore they will elaberate a bit more for you |
Matt Send message Joined: 6 Jun 10 Posts: 18 Credit: 1,434,540 RAC: 0 |
Thank you Glenn for you response. I've been in and out of S@H for a few years and I'm not sure what Lunatics is, so, yes, I run the stock BOINC app, though with an app_config which allows me to run multiple WUs per GPU. For anyone looking for details on my rigs, here is my GTX 680 machine and my GTX 780Ti machine. |
Darth Beaver Send message Joined: 20 Aug 99 Posts: 6728 Credit: 21,443,075 RAC: 3 |
lunatic is also known as anonymous platform. your GTX 680 machine had at some point a ATI GPU ? yes your GTX 780 is doing mainly cuda 42 I would suggest you run Lunatic installer and use the cuda 50 and multibeam Astropluse . you should be able to run 4 units on that 780 as it has 2 cuda gpu's and the GTX 680 shouold also be able to do 4 units as it also has 2 cuda GPU's but you will have to reserve at least 2 cores on each machine and possibly as many as 4 cores to be able to feed the GPU prity good machine you got there a real cruncher |
Darth Beaver Send message Joined: 20 Aug 99 Posts: 6728 Credit: 21,443,075 RAC: 3 |
oh sorry i forgot after running Lunatics you will have to change the app-inf.xml file to do multiple units on the gpu s it will have as the default i unit per gpu. But with Lunatics they will only use the Cuda version you specify and will run faster than stock Boinic |
Matt Send message Joined: 6 Jun 10 Posts: 18 Credit: 1,434,540 RAC: 0 |
Thanks, Glenn. Yes, My 680 machine was originally an all AMD box - first with a 6950, then 6950 + 7950, then 7870 (Tahiti) + 7950, and now two GTX 680s w/ AMD Phenom II 965 CPU. The Lunatic installer is something I should Google and install? I'm really not sure what it is or what advantages it offers over the stock app via BOINC. |
Darth Beaver Send message Joined: 20 Aug 99 Posts: 6728 Credit: 21,443,075 RAC: 3 |
no you don't have to google it Arkan i believe has the installer on his web site give a few minutes i'll find the link and post it for you There is a big advantage using Lunatics it should increase the crunching by 50% but not only will it increase the speed on the GPU units it will all so increase the speed of units done by the CPU |
Darth Beaver Send message Joined: 20 Aug 99 Posts: 6728 Credit: 21,443,075 RAC: 3 |
|
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
All the wus/tasks are the same. There is no such things as a "CUDA task" of any version. The "version" is determined by the app that was assigned when the work was requested. When using the stock applications the server sends each app. Then once it determines which runs best on your host it will then run it exclusively or most of the time. It may check that the other versions at a given interval. I don't recall. With the optimized apps you can specify the CUDA version at the time of the installation. However, when using the optimized apps you have to update the apps manually when they change. IIRC app_config.xml takes precedence over app_info.xml. So you may not have to modify anything if you choose to run the optimized applications. If you are going to get the optimized apps. Once you get to the Lunatics downloads. Looks for the line They are available from arkayn's Crunchers Anonymous Windows downloads and Mike's download page.. Then on either site get the Lunatics v0.41 setup for the platform you need. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
arkayn Send message Joined: 14 May 99 Posts: 4438 Credit: 55,006,323 RAC: 0 |
ok here is Arkan's site That is not my site, mine is http://www.arkayn.us/forum/index.php |
Josef W. Segur Send message Joined: 30 Oct 99 Posts: 4504 Credit: 1,414,761 RAC: 0 |
Hello Everyone, BOINC attempts to characterize the effective speed of each app version on your host, that measurement is shown on the application details page for each host as "Average processing rate"(APR). When the host does a work request, the servers choose the best app version primarily based on the APR. However, there's a random factor introduced so app versions with lower APR do get some work. The goal is that even if the APRs have been skewed by variations in work characteristics or other uses of the computer the choice of app version will become more reliable over time. The amount of randomization is reduced as more entries go into the averages. Application details for your two GTX 780Ti rig show that the mix of work has made the CUDA42 app version's APR much higher than that for CUDA50 and CUDA32, hence it is much more likely to get CUDA42. The "Number of tasks completed 648" for CUDA42 means not much randomization will be applied to its APR, but the basic random factor used has a normal rather than equally weighted distribution so even that ~350 GFLOPS will occasionally be seen as much different. The lower counts for CUDA32 and CUDA50 mean more randomization. For your two GTX 680 rig, the APRs are closer and all app versions have lower counts so the randomization is still causing a more even mix. Joe |
Matt Send message Joined: 6 Jun 10 Posts: 18 Credit: 1,434,540 RAC: 0 |
Thanks everyone for taking the time to post. I learned a few things and will definitely check out Arkayn's site for Lunatics. I didn't realize the Application Details page existed for my hosts. That would have answered a lot of my questions up front. |
Darth Beaver Send message Joined: 20 Aug 99 Posts: 6728 Credit: 21,443,075 RAC: 3 |
Sorry Arkan i did go to your site and then i clicked on your link to Lunatics probbly should have said that |
Bill Greene Send message Joined: 3 Jul 99 Posts: 80 Credit: 116,047,529 RAC: 61 |
Does anyone have a simple algorithm for determining the number of WU's to assign a GPU? I'm bringing online a new system with 2 X 780ti's and would like to avoid the trial and error approach if possible. Presently I'm running 2 X 480's with 3 WU's assigned each. Seems to execute timely but for some reason I'm receiving 5-10 invalid results daily on that machine; from 10-20 per day have status of 'validation inclusive'. CPU cores have not been set aside beyond the automatic reserves. And of course I'm generating interrupts from other uses. Sure would like to clean up results from the 480 machine as well as take advantage of experience on use of 780's. |
Zalster Send message Joined: 27 May 99 Posts: 5517 Credit: 528,817,460 RAC: 242 |
We would need more info on your set up. MB, make of chip and cores, etc. Depending on what is supporting the GPUs will determine how best to go about determining how much work load they can do. |
The_Matrix Send message Joined: 17 Nov 03 Posts: 414 Credit: 5,827,850 RAC: 0 |
@Bill Just use the freeware GPU-Z , and look about the usage of a GPU under sensors. 50% usage = 0.5 gpus |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13736 Credit: 208,696,464 RAC: 304 |
Presently I'm running 2 X 480's with 3 WU's assigned each. I can't see how that would be causing errors, unless your power supply is running right at the ragged edge. Grant Darwin NT |
Bill Greene Send message Joined: 3 Jul 99 Posts: 80 Credit: 116,047,529 RAC: 61 |
Appreciate all responses. I'll address each in this posting - By Zalster: We would need more info on your set up. MB, make of chip and cores, etc. Depending on what is supporting the GPUs will determine how best to go about determining how much work load they can do. Configuration consists of: MB ASUS|X79 DELUXE R VGA EVGA|03G-P4-2884-KR GTX780TI 3G X 2 PSU EVGA | SUPERNOVA1300G2 1300W RT CPU INTEL|CORE I7 4930K 3.4G 12M R SSD 500G|SAMSUNG MZ-7TE500BW R MEM 4Gx4|G.SKIL F3-12800CL9Q-16GBXL MS WIN 7 HOME SP1 64BIT SLM - OEM WATER COOL CM | RL-S24M-24PK-R1 R CASE CM|HAF X BLUE ED RC-942-KKN3 R and a 2TB HD I have on hand. By The_Matrix: Just use the freeware GPU-Z , and look about the usage of a GPU under sensors. 50% usage = 0.5 gpus I'll check this out further. Initial read (at download.com) indicates "GPU-Z is a basic tool that launches a small, tabbed interface with no bells and whistles, just information in a dialog box-style window". Free executables tend to raise the hair on the back of my neck - trojan horse, etc - but I'll give it a good scan before using. Thanks for the lead. By Grant: I can't see how that would be causing errors, unless your power supply is running right at the ragged edge. Presently using about 735 watts on a 1050 watt PSU. This dual 480 machine ran 3 WU's each without errors until I installed a system SSD. However, when I went back to the HD, the errors continued so I'm back on the SSD (a surprising approximate 5% gain for the SSD even with errors). I've reloaded Lunatics to no advantage. May have to do some trial and error cycles with operating parameters but presently have no suspects. As before, appreciate any thoughts ... |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13736 Credit: 208,696,464 RAC: 304 |
Probably worth posting the Stderr_output for a few of the invalid ones so those that know can have a look at them. Validation inconclusives aren't an issue, unless they turn out to be an error. My inconclusives tend to be between 30-50, those with faster systems will see much higher numbers of inconclusives. Another programme in addition to GPUz would be DPC Latency checker. Large spikes in DPC latency can cause all sorts of very difficult to track down & resolve issues. I'm running 2 GTX 750Tis on a an i7 2600K with Win7 and 2 GTX750Tis on a C2D E6600 with Vista, both systems using v335.23 video drivers. The only errors I get are truncated Stderr_outputs about once a month or so. Grant Darwin NT |
Zalster Send message Joined: 27 May 99 Posts: 5517 Credit: 528,817,460 RAC: 242 |
Well, I don't run as powerful a CPU chip as you have (I have a AMD) or have the SSD. I run 2 APs on the 780s when available then 3 MBs on them when we are out. No CPU work at all. Are you running any work units on the remaining CPU cores when you crunch. Is Hyperthread engaged on your system? I seem to remember someone talking about HT and the debate about whether it should or should not be used. I have no opinion on this since I don't have an Intel chip (other than the store bought one at home in the den). Other than that, why would the adding the SSD drive start to cause you to have errors? That is interesting. I'm assuming no power saving modes and no spinning down of the HD after a certain amount of time via the preference panel. On those work units that errored? It was after they were compared to your wingman or did they error right away? If they errored right away there might be a clue in the stdeer report. Let see if anyone else has any ideas. Zalster |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.