I've got a 1060 6GB in an Akitio Node Thunderbolt 3 enclosure, and during the night, I'd like it to help heat my room, so I use it for BOINC.

I need to prevent GPUGrid from trying to use my laptop's built-in 940MX, though.

So, I searched around, and came up with cc_config.xml options of <exclude_gpu> <device_num> 0. That worked to avoid the 940MX. However, like most modern laptops, the computer also has an Intel GPU... I knew that'd get a bit tricky.

Seems that GPU numbering in BOINC is not well documented nor consistent - but worse than that, GPUGrid doesn't even honor the command line settings it's given.

If I run acemd-922-80.exe with --device 0, it reports back the 1060 (that's not right in any case, except maybe if that means "give me default" and nVidia funnels it to the fastest GPU).
If I run it with --device 1, it reports the 940MX. Sadly, that's the command line that BOINC is giving it, while believing it's running on "device 1" (my 1060).
If I run it with --device 2, it runs again on the 1060.

BOINC thinks that device 0 is the 940MX and device 1 is the 1060, and the Intel GPU is its own special little thing.

So, BOINC is telling GPUGrid to run on device 1 (1060) but GPUGrid starts running on the 940MX instead. Basically, to use GPUGrid, I'll have to dedicate my machine to doing only that, because BOINC and GPUGrid will overlap WUs on a single GPU instead of using one WU one each.

Le halp?

With multiple manufactures they will each be device 0. You'll probably need the <type> command as well to exclude the Intel device at index 0.
https://boinc.berkeley.edu/wiki/Client_configuration

The event log at start of the BOINC client will mention the Device ID per manufacture.

Hm, but the problem isn't that it's trying to run on the Intel card, but just that the index is different between BOINC (which calls GPUGrid), and GPUGrid itself. So BOINC thinks it's telling GPUGrid to run on one card, but it's really running on another.

Excluding a GPU from BOINC will just keep a WU from running on a device, but it won't change the indexing of cards.

Present example - a lucky break - is that two GPUGrid WUs are running right now - one "short run" and one "long run". It thinks that the "long" is running on device 0 (the 940MX) and the "short" is on device 1 (1060). The reality is, thankfully, the opposite. But this issue prevents me from keeping GPUGrid from running on the 940MX at all, which puts WUs at risk of being aborted because the 940MX is too slow for long run WUs... and I'd much rather have the 940MX crunching other projects instead.

Just curious, what does the startup log says about all this? and what does your cc_config look like. Could you post both so we can look them over?

Sure - here's my cc_config.xml, just enabling multiple GPUs:

I see a couple of issues.

First your app_config is missing a <app_config> somewhere in there as BOINC is ignoring it.

Looks like BOINC assigns the 1060 as device 0 when it's connected and when it's not then the 940MX is device 0

We can ignore the intel build in as it does not have CUDA so none of the work units will run on it.

So that takes us back to only 2 usable GPUs, the external 1060 and the built in 940MX.

You will probably have to use the ignore device 1 to use the external GPU but have the device attached prior to starting BOINC. If you wait to attach after you launch BOINC then it's going to get confused.

Let's see if straightens things out.
____________

BOINC is seeing 3 GPUs, two of which are device 0. 1 per device manufacture. Setup cc_config per the Wiki link with device 1 for the 940MX with type option.

The intel GPU is not supported by GPUGrid so I don't think this will affect the GPU order

Right, the Intel GPU was never part of the problem :P

My app_config is blank, intentionally, as I originally tried setting the options there (where I wanted it to be app-specific) and just left the blank file there instead of digging in to delete it. Shouldn't be affecting anything.

My goal is to be able to use BOINC with or without the 1060 attached, without it trying to run GPUGrid on the built-in GPU. So, making sure the 1060 is attached first isn't much an option (plus, BOINC can't hotplug GPUs, so it has to be shut down first to attach/remove the GPU).

I'll have to look into that "type option" real quick. That'd do the trick if it's a thing :)

edit: there doesn't appear to be any way to direct an ID assignment to a specific card name with cc_config.xml. :(

also overnight update: 8-1/2 hours into crunching, and the 1060 is 78.5% done with the Long Runs WU, and the 940MX is 35% done with a Short WU. Yeah, kinda expected that... imagine if it accidentally got a "long run" ;)

Ok, now that I have a better understanding of what you want to do, we should try to BOINC to ignore the built in GPU. Never tried this on GPUGrid but.....

On Einstein I did have success with the following

That would just keep GPUGrid from using any device numbered "1", which would keep the 940MX from doing anything when the 1060 is connected (as 0), but with it disconnected (as now, writing this from bed :) ), it'd revert to using device 0, which would be the 940MX.

I guess there's just no way to do this with the current BOINC architecture, without a way to identify a specific GPU by name. Not a huge issue now that I know the device IDs are properly assigned, though I worry if a "long" WU gets stuck being computed by "1". During the cold winters, I'm almost always running BOINC with the GPU connected for heating, so it's OK.

Probably best to add that just as a safeguard, if nothing else. :)

Thanks for the input, everyone! I think it was a mix of my own brainfart (in the initial subject, here) and limitations of BOINC itself. I'll call this resolved!

That would just keep GPUGrid from using any device numbered "1", which would keep the 940MX from doing anything when the 1060 is connected (as 0), but with it disconnected (as now, writing this from bed :) ), it'd revert to using device 0, which would be the 940MX.

I guess there's just no way to do this with the current BOINC architecture, without a way to identify a specific GPU by name. Not a huge issue now that I know the device IDs are properly assigned, though I worry if a "long" WU gets stuck being computed by "1". During the cold winters, I'm almost always running BOINC with the GPU connected for heating, so it's OK.

Probably best to add that just as a safeguard, if nothing else. :)

Thanks for the input, everyone! I think it was a mix of my own brainfart (in the initial subject, here) and limitations of BOINC itself. I'll call this resolved!

That's the thing, there are two nVidia cards here when the 1060 is connected. It shifts so that 0 = 1060, 1 = 940MX (BOTH are nVidia!), but when the 1060 is unplugged, it becomes 0 = 940MX (again, nVidia). So all I'd be doing with that config is excluding "whatever device 1 happens to be", which would still leave it with 0 = 940MX being enabled when the 1060 is unplugged.

It's OK. I did that anyway, since I'm not really running BOINC with the GPU unplugged. It just needs to keep the 940MX from trying to crunch a "long" WU ;) And it's doing that just fine now. :)

I'm going to be going this route myself and exclude the RTX 2080 from both Einstein and GPUGRid. Want to keep the other 3 Nvidia cards running those projects though.

Glad I could help Keith.
____________

I'm not having any luck with excluding device 0 for both GPUGrid and Einstein. Doing so prevents all Seti cpu tasks from running and puts them into "waiting to run" status.

The excludes are recognized as correct in the Event Log startup. But they prevent Seti cpu tasks from running. Seti gpu tasks continue to run as well as Milkyway gpu tasks.

Don't know what's causing the problem. Only way to make things work correctly for Seti is to remove the excludes for a basic cc_config.

For now I will just have to suspend both projects on that host until someone can tell me what is going on and how to correctly exclude the 2080 from those projects but allow the other Nvidia cards to run those projects.

Keith PM your cc_config so I can see what it looks like.
____________

Okay, after a while of crunching like this, I've determined that there IS definitely a bug of some kind in how GPUGrid and/or BOINC handles device ID assignments.

Take a look at this screenshot:
https://imgur.com/1BjG4ka
(can't seem to attach, and also can't seem to thumbnail-link, so this is the best I can give you. jeez, I even had to URL-tag wrap it to make it clickable)

Here, you can see BOINC is assigning GPUGrid to ID#0 (the 1060) and Einstein@Home to ID#1 (which should be the 940MX) - a perfect assignment.

However, if you look at the nVidia status window, it shows both projects competing for processing power on the 1060. It works, but it dramatically slows down the GPUGrid workunit. I discovered this just last night as a WU I was crunching had been running for over 1 full day on the 1060, something that should obviously never happen... as soon as I was able to rip other projects away from that GPU and let the same WU continue running, its GPU TDP shot up by about 30% (from 40% TDP to 70% TDP) and the GPUGrid WU started crunching way faster. Still not sure if it completed the WU in the deadline... :(

I tried just shutting off GPU computing for all other projects, as so:

wow this forum's functionality is extremely limited (not being able to edit after a time period, not being able to post images inline without an external host, etc etc etc 2003 internet things)

anyway, solved my own problem, found that nVidia's own "what's running on what" indicator is actually faulty. Another dual-GPU system I had was reporting similar issues with BOINC, but I wasn't even using GPUGrid on that one. It's just the tool that's stupid. GPU-Z shows the real info, and all is well in the world.

Since you are using windows, might want to look at BoincTasks and SIVx64. I like it as it gives better idea of what my system is doing.

http://rh-software.com/
https://efmer.com/
____________

Have you tried Nvidia's built-in tool that is available? nvidia-smi?? It shows what is running on each gpu, the amount of utilization, memory usage, power usage in watts and each application running on each particular gpu.

You can find it in C:\Program Files\NVIDIA Corporation\NVSMI

Just open a command window and type nvidia-smi.

Have you tried???