Questions and Answers :
GPU applications :
How to use multiple GPU's
Message board moderation
Previous · 1 · 2 · 3 · 4 · Next
Author | Message |
---|---|
Bravo6 Send message Joined: 28 Oct 99 Posts: 52 Credit: 44,947,756 RAC: 0 |
I think I have covered most of the bases, thinking about this project, but one last question popped up before starting. Given the example of the gtx 1050 series, is it sufficient to match model and and memory size, and what would be the effect of straying too far in matching the cards in a system? Thanx "Don't worry about it, nothing is gonna be O.K. anyway.........." |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
I think I have covered most of the bases, thinking about this project, but one last question popped up before starting. Given the example of the gtx 1050 series, is it sufficient to match model and and memory size, and what would be the effect of straying too far in matching the cards in a system? One effect of having different cards in one system is it is harder to run the higher end cards at full speed because unless you can do some fancy tinkering (that I don't know how to do) you will end up using the same command line on all cards of the same brand (eg. NVidia and AMD). And if you can run 3 tasks on say your Gtx 1080Ti but you also have a Gtx 1050Ti that you can't run three tasks on, you would have to see if you could figure out a way to give each card a different "app_config.xml" parameter. Which is something else I don't know how to do. So in general, running multiples of the exact same card even if they are different brands is easier (at least for me). Tom A proud member of the OFA (Old Farts Association). |
Jord Send message Joined: 9 Jun 99 Posts: 15184 Credit: 4,362,181 RAC: 3 |
...you would have to see if you could figure out a way to give each card a different "app_config.xml" parameter. Which is something else I don't know how to do.Run multiple clients. But that's Astrophysics level BOINC. |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
...you would have to see if you could figure out a way to give each card a different "app_config.xml" parameter. Which is something else I don't know how to do.Run multiple clients. But that's Astrophysics level BOINC. Thank you for the reminder. I had forgotten. Tom A proud member of the OFA (Old Farts Association). |
Bravo6 Send message Joined: 28 Oct 99 Posts: 52 Credit: 44,947,756 RAC: 0 |
Ok makes sense. While I used to install and manage multiple server level systems, it is amazing to me the change/advances since the" dark ages " (for me that's 2005). I am afraid my level of expertise has degenerated to "not too dumb, considering". Now we will see............ Thanx..... 😨 "Don't worry about it, nothing is gonna be O.K. anyway.........." |
rob smith Send message Joined: 7 Mar 03 Posts: 22527 Credit: 416,307,556 RAC: 380 |
A couple of tips if you are contemplating running mixed AMD & nVidia GPUS - get each type stable and working independently and save the configuration files before trying to get the other one working. For each type start with one GPU the build up to the "final fleet" size". Then when you start to think seriously about the mixed environment start with a small count and build up. From memory there is flag in BOINC that needs to be set to allow multiple instances to run, which you will need to do for a successful mixed fleet. There has been some discussion in the "Number Crunchers" forum inside SETI (https://setiathome.berkeley.edu/forum_forum.php?id=10) that covers this sort of activity. Take your time to get each step right before moving on to the next, and keep copious notes of what you've done so you can roll back when you get things wrong or want to try something slightly different. Good Luck Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Bravo6 Send message Joined: 28 Oct 99 Posts: 52 Credit: 44,947,756 RAC: 0 |
I wasn't thinking so much of different cards so much as different versions of same card. For example the gtx 1050s, and 1050to with different men size, different DirectX , versions and the other little things different for production run to production run. I Am going to start with two Hewlett packard refurbs, one I3 and one I5 and use 1050s in one and 1050ti in the other. ( i have access to several of each). Just waiting on delivery from #$%&%$ china. Did not realize that was where stuff i ordered was going to come from. I like newegg , but ya gotta read the source print........ 😒 "Don't worry about it, nothing is gonna be O.K. anyway.........." |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
I wasn't thinking so much of different cards so much as different versions of same card. For example the gtx 1050s, and 1050to with different men size, different DirectX , versions and the other little things different for production run to production run. I Am going to start with two Hewlett packard refurbs, one I3 and one I5 and use 1050s in one and 1050ti in the other. ( i have access to several of each). Just waiting on delivery from #$%&%$ china. Did not realize that was where stuff i ordered was going to come from. I like newegg , but ya gotta read the source print........ My experience with various brands of the same models of Nvidia (mostly Gtx 750Ti's and 3 Gtx 1060 3GB) is that you can use the same parameters for the same models even if the performance is different. I have had two 750's and two 1060's running under Windows where I could use the GPU-Z utility on them. Inspite of being the same model cards there was differences in speed and some differences in productivity. My Gtx 750Ti's have (I think) 2 GB of memory. And the 1050's apparently start at 2GB and go up. I hope you ordered "brand name" cards (eg. ASUS, EGA etc.) not just "generic" Nvidia cards. My experience with "generic" Nvidia cards even from the US is they are not the actual product you thought you were getting. Tom A proud member of the OFA (Old Farts Association). |
Zalster Send message Joined: 27 May 99 Posts: 5517 Credit: 528,817,460 RAC: 242 |
I wasn't thinking so much of different cards so much as different versions of same card. For example the gtx 1050s, and 1050to with different men size, different DirectX , versions and the other little things different for production run to production run. I Am going to start with two Hewlett packard refurbs, one I3 and one I5 and use 1050s in one and 1050ti in the other. ( i have access to several of each). Just waiting on delivery from #$%&%$ china. Did not realize that was where stuff i ordered was going to come from. I like newegg , but ya gotta read the source print........ You are always going to be limited by the lowest card. That mean the one that is the slowest or has the least amount of Ram. To that end, your commandline and setting will always be dictated by that card. Example. If you had a 1050 and a 1070, you would be restricted by the 1050. If you had 2 1050s but one had 2 GB of Memory and the other 6GB, you will not be able to craft your commandline to take advantage of the extra memory. So it's best to always have similar cards. If you chose to use different series of cards, like a 1050 and 1070, then you will be forced to use a cc_config.xml to force bonic to use both cards. Otherwise it will ignore the 1050 and only use the 1070. Lastly, there is a limit on how much RAM of a card you can use for OpenCl. Not such a problem with cuda but depending on which application you choose to run, it will also limit how you can tailor your system. |
Bravo6 Send message Joined: 28 Oct 99 Posts: 52 Credit: 44,947,756 RAC: 0 |
ok added a gtx 1050ti to one on my newer refurbs got it all recognized yada yada and the work units i had waiting dissappeared and i am back to only getting the work units the system can handle at one time. I understand (i thjink) that the system will start letting this machine download up to a hundred wu s after a while, but this machine had about a hundred wu s waiting to run. Did adding a card the a system already getting many work units downloaded put it status back and orphan the wu's already waiting? I am working thru adding multiple gpu's to several systems, and the new card here may be temporarily removed while building out. Am i going to have more problems like this? and am i causing problems for the larger system? ShouId i get the system i am adding to down to not having workunits actively downloaded and then add the gpu? don t want to mess up beacoup wu's while "messing around" I guess i m still ahving problems, i just noticed that after running a few wu's, the new gpu is now not seen by seti pgm...... sound familiar to anyone? lokks like lotsa fun coming up....... Just re-ran the cpu benchmarks and the new cpu showed up again......... will watch it..... looks like I ll need lots of notes!! thanx "Don't worry about it, nothing is gonna be O.K. anyway.........." |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
For your A10 you have a special case that doesn't apply directly to your Intel cpus as long as you are not using their internal gpu's to process seti tasks. I would either shutdown the AMD gpu tasks or throttle the cpus to 75% leaving 1 "clear". Either of these will speed up the whole system processing. If you turn off the AMD gpu tasks, they will have to "clear out" or be aborted before you will see any improvement. If you use the local configuration to reduce the # of cpus from 100% to 75% that will drop out one cpu, allowing the internal gpu to process at "full speed" while the other 3 cpu's also process at full speed. I don't have any empirical evidence about how the GTX 1050Ti will run if you throttle the cpu's. But I am hopeful it will run at "full speed". HTH, Tom A proud member of the OFA (Old Farts Association). |
Bravo6 Send message Joined: 28 Oct 99 Posts: 52 Credit: 44,947,756 RAC: 0 |
Don t know much about the throttling options, but the AMD solved the problem temp by dying, I think it is either a glitch in the power supply or a dead fan. It is a cyberpower trinity, so I have to do some deconstructing....😞 I m not sure that I understand what you are saying about throttling the cp u s down from 100% to 75%? Where can I do that? Is there somewhere a decent tutorial on the deeper options of the seti software, I tend to like getting in the weeds when I have spare time, but haven't found a documented source yet, is there one around? Also I am assuming that the computing options are individual per computer and not "groupwise" is that correct? I am using the local config files (I think). The more I read the more complexities i am seeing..... Thanx "Don't worry about it, nothing is gonna be O.K. anyway.........." |
rob smith Send message Joined: 7 Mar 03 Posts: 22527 Credit: 416,307,556 RAC: 380 |
Last question first... It all depends where you set your "computer options", and how you've got other flags set. If you set the computer options on the website, and do not have the "use local prefs" flag set then settings work at "group" level. If you do them using the various configuration files on one computer then then they always only apply to that PC. If you set them using the BOINC manager interface then they can only apply to the local computer, but, in many cases only if the "use local prefs" flag is set. Throttling of a CPU or GPU reduces the amount of work it is able to do, thus its running temperature. There are various ways of doing it (didn't you guess?). There are tools like "Tthrottle" that allow you to control clock speeds and the like, but they aren't normally needed unless your cool is very marginal, or you are seriously overclocking (so why not just down clock a bit and reduce the temps - that's all these tools do). Or you can tell BOINC not to use all the CPU cores (that can be set using the BOINC manager advanced view) - simple and immediate. Or in the case of an A10 you can opt not to use the APU for processing. People have provided "mixed reviews" of the performance of the APU in the Ax family from AMD - the A10 is the top of the tree, so has quite a fair performance, certainly better than a CPU core, but not as good as a GTX750. It will need a bit of care when it comes to setting up the drivers as AMD GPUs (and by inference APUs) are "very fussy" over driver setting compared to the nVidia family. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Bravo6 Send message Joined: 28 Oct 99 Posts: 52 Credit: 44,947,756 RAC: 0 |
O. K, a big "duh" is due from me. My brain (unique in its own right) wrapped around those options wrongly and was thinking "time" not "horsepower". Should have known better but was probably in a hurry. Have to start looking for documentation on the more arcane options....... Want to make my setup as efficient as my poor brain allows...... Thanx "Don't worry about it, nothing is gonna be O.K. anyway.........." |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
Bravo6, My discussion about dropping 1 cpu off the A10 was based on some extensive testing done on another website (I think I read this on someplace on Lunatics where petri got very serious and took lots of notes). It also is something that is specific to the AMD "All in One" aka: A-x to A-10. I believe that the "choking" problem that I have experienced using the A-series is specific to that design and should not be generalized to all systems, either AMD or Intel. As Bob mentioned, a way to reduce the number of cores/threads being run on a cpu is to use the individial configuration options that are available in the Boinc manager. Since you appear to be running only one A-10, I would go that route rather than setting it "group wide" from the website. I understand something about the A-10 has died. Hope it isn't the motherboard/cpu :( I expect trying to boot it without a video card and with a known good PSU will answer the question of if the MB has bit the dust. I had a lot of trouble with my A-10. Its driver regularly crashed and when it restarted, the Seti gpu processing didn't restart. I would go with the external card and forget the internal one. Good luck! Tom A proud member of the OFA (Old Farts Association). |
Darrell Wilcox Send message Joined: 11 Nov 99 Posts: 303 Credit: 180,954,940 RAC: 118 |
@Zalster I run a mixed GTX750Ti/GTX1070 in a couple of systems, and have found at least some of the parameters can be specified for the LARGER GPU, and the smaller GPU uses its maximum. For example, I specify SBS 2048, and the GTX750Ti says Used GPU device parameters are: Number of compute units: 5 Single buffer allocation size: 512MB Total device global memory: 2048MB while the GTX1070 says Used GPU device parameters are: Number of compute units: 15 Single buffer allocation size: 1024MB Total device global memory: 8192MB This is with the SOG (stock) version. I haven't messed with other variables so if you do, let the community know how it goes. |
Kissagogo27 Send message Joined: 6 Nov 99 Posts: 716 Credit: 8,032,827 RAC: 62 |
some Raistmer research about Igpu limiting cpu computation http://lunatics.kwsn.info/index.php?topic=1735.0 |
Tom M Send message Joined: 28 Nov 02 Posts: 5126 Credit: 276,046,078 RAC: 462 |
some Raistmer research about Igpu limiting cpu computation Thank you. I was looking for that URL! Tom A proud member of the OFA (Old Farts Association). |
Kissagogo27 Send message Joined: 6 Nov 99 Posts: 716 Credit: 8,032,827 RAC: 62 |
;o) |
Zalster Send message Joined: 27 May 99 Posts: 5517 Credit: 528,817,460 RAC: 242 |
@Zalster That's fine but you have to understand how memory is set by Nvidia. Only 26% of total ram is available for OpenCl use. Period _iterations is normally 50. I would not decrease that value as the 750 might have problems. Here's an example of a commandline. -sbs 1024 -period_iterations_num 10 -spike_fft_thresh 4096 -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64 The ones in yellow are all you really need. All the rest can be used but that is for a high end GPUs. I never used them on my cards. I would stay away from -hp since there is a 750 in that case. |
©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.