I agree with skgiven -- Your GPUs are getting too hot. For a GTX 780, althought it is rated to support up-to-100*C, that is it's "absolute maximum operating temperature" (TjMax). It actually starts thermal downclocking at 80*C, since GPU Boost v2.0 GPUs use an 80*C threshold. And 80*C is generally about the "maximum comfort zone" for a GPU that you want to take care of.

I'd investigate your cooling, and then consider using a program like Precision X or Afterburner to enforce a custom fan curve that keeps it below 80*C.

Well, I have good news. If the current gold version of the BOINC client does have a bug to not use all 3 GPUs, the .36 version does not. I updated the driver and the BOINC and now it is crunching n 3 GPUs with 3 tasks!

Funny that this whole conversation could have been solved (but I would not have learned as much) if someone just gave me the link to 7.4.36 and said, "Here, try this." lol

I will check on that heat thing.

Thank you thank you thank you so much for all your help. (I will not close the door on this until I see this repeated, but I will operate as if it is.)

Both your GPUs 0 & 1 are hitting 95-96C I wouldn't take too long to try to solve the temp problem.

OK done, using Afterburner.

Glad to hear you got things working and your priorities sorted out, hopefully.
The 780's are probably quite difficult to keep cool (and quiet), but it's well worth it.

- Make sure you set a fan profile for each/all of the cards.

PS. Boinc not seeing the GPU was probably down to the driver or card swapping:
In layman's terms it's not Boinc's fault as Boinc only reads what the drivers tell it. The ACEMD app doesn't ask Boinc, it looks & reports, but as GPUGrid uses Boinc it can still only use the resources Boinc says are available!

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

This seems like a new topic, but it also seems like something those already responding here can answer for me. Sorry if this is answered all over the forums and for my own lack of investigation. I could not find anything with a few searches of the terms I was looking for.

I don't quite understand how the GPU works compared to a CPU with shaders and all that goes with them. With that in mind, I found that my laptop has a mobile version of an NVidia card and was able to run GPUGrid on it. After it started running, I noticed that it is running at 78 degrees, which is fine, but it is running at 97% GPU load also. This struck me as odd since the GPU loads on the 780s in the big rig run at loads in the mid to high 70s. Since my last post, I found a new cooling solution for that rig (it being super cold out now) with a window open and a room fan aimed at it. And since that is keeping the temps in check without limiting with Afterburner and knowing overclocking a GPU will corrupt GPUGrid results, I want to know how to push the GPUs to use closer to 100% load. I do have it running cooler and can increase the cooling to make it run even cooler yet, if need be.

1) The same GPU load is reported by Afterburner, OpenHardwareMonitor, and NVidiaInspector, but is that reported load correct according to actual load or is it just an semi-accurate approximate?
2) Is the reported load of 7x% actually close to 100% load but with things I don't understand (like shaders, VPU, MCU, Bus speed, Memory speed, "Boost", etc.), it is showing me a number that looks much lower?
3) Is the GPU load a task specific item? I only ask this because I do see some tasks in the 60s and some in the 80s.
4) If it really is running in the 70s and 80s, how can I get it to run the full GPU load?

If you would like to answer simple answers, I can take that or if you can give or point me to definitive detailed answers, I would appreciate those.

Thanks again,
Mike

---

1 Corinthians 9:16 "For though I preach the gospel, I have nothing to glory of: for necessity is laid upon me; yea, woe is unto me, if I preach not the gospel!"
Ephesians 6:18-20, please ;-)
http://tbc-pa.org

Mike, there are four main things to consider here; the system architecture (especially GPU), the WU/app, exactly what is GPU load/utilization/usage and what else you are doing on the system.

Your systems are both of high spec, so there is no obvious bottlenecks or performance differences that would impact GPU usage.
The Quadro K2100M is a GK106 with a 576:48:16 core config (Unified Shaders : Texture mapping units : Render output units).
It's often the case that a smaller core results in greater GPU utilization (according to the apps/tools you mentioned).
Compared to your 780 (2304:192:48), your K2100M is basically 1/4 the size and has a slightly higher Render Output Unit ratio, but it also has a 128bit bus compared to the 780's 384 thats relatively high (1/3rd) and I expect it results in a reduced memory controller load for the Quadro.

GPUGrid factors that change GPU utilization are the Work Unit type; some result in a higher GPU utilization than others. Incidentally, the WU's that utilise the GPU less tend to use less power. It's likely that these WU's simply don't require to utilize the GPU as much. Different WU types use different amounts of GDDR, different amounts of the CPU and impact on the Memory Controller Load differently too. Only the researchers could tell you exactly how each WU utilizes your GPU (especially the shaders) differently and why some tasks use more CPU, GDDR or bandwidth, but it stems from the molecules being examined; the number of atoms.
Note that the relative performance of different cards varies with the WU type, making it more difficult to compare GPU's. Different apps can also impact performance, though in your case the apps are the same.

Exactly what is being measured to determine GPU utilization by the tools is a good question, but not one I can answer. I know that the NVidia facilitate this, so perhaps they would be the ones to ask.

If you are using a system heavily it will negate performance somewhat. If you are trying to run 10 CPU apps and 3 GPU apps for example the GPU performance will drop off a bit.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

In order for me to get better GPU utilization added a file in C:\ProgramData\BOINC\projects\www.gpugrid.net called app_config.xml with. I have 3 GPU's in the machine. This lets you run two WU's per GPU with each getting half. My GPU utilization runs around 98% on each GPU.


<?xml version="1.0"?>

-<app_config>


-<app>

<name>acemdlong</name>

<max_concurrent>6</max_concurrent>


-<gpu_versions>

<gpu_usage>.5</gpu_usage>

<cpu_usage>1</cpu_usage>

</gpu_versions>

</app>

</app_config>

I want to know how to push the GPUs to use closer to 100% load. I do have it running cooler and can increase the cooling to make it run even cooler yet, if need be.

Thanks again,
Mike

A simple app_config.xml file will let your run multiple gpu units on one cpu at once, thereby utilizing your gpu to it's max. The problem will come in that since you are then pushing your gpu to work harder, the units will EACH take longer and you may not get an 'bonus' credits for finishing the units within the shorter times you are now. A unit is using about 70% of your gpu right now, that means thee is not enough room to load a full 2nd unit so it will have to be sharing some of the gpu's resources to run, slowing down each one. At most projects that isn't a problem as there are no 'bonus' credits for finishing units faster, but here there are. I did not even address the heat issue of pushing your gpu harder!

All this comes down to you have a gpu that has more capability than what the programmers designed their software to run on, and it is just cruising thru the units, while the rest of us with our older gpu's are struggling. You are at the tippy top of the spear right now, in a few years, when the rest of us upgrade to something even better, we will pass you on by and you will be the one struggling, enjoy your time out front while you can, it will end! I for one am envious, but shopping!

Is the GPU load a task specific item? I only ask this because I do see some tasks in the 60s and some in the 80s.

Yes. It depends on:

- The speed of your GPU. The faster it is, the lower its load (since there are always small pauses where CPU support is needed)

- The number of atoms being simulated, i.e. the complexity of the WU. This can be seen in the task output in your profile.

- The physical model choose by the scientist. The more work the GPU has to do before it needs CPU support again, the less pauses occur per second.

Edit regarding running multiple concurrent WUs: if you're running at 85% GPU load or better, there's little benefit for even a performance loss from doing it. Below 80% load throughput improves. Those numbers are not exact, but the turning point is somewhere between them.

MrS

---

Scanning for our furry friends since Jan 2002

I found a way to get over 90% of the GPUs working. Since Dnetc is not a BOINC project, I set the Dnetc GPU priority to 2 and it took up the slack from GPUGrid. Unfortunately there is no happy medium and it is slowing down the GPUGrid project, but my BOINC stats went from 1,500,000 to 850,000 a day round abouts, but my Dnetc stats went from 7,000 a day to over 50,000. I wish I could get some happy place where the DNetc only took the idle that GPUGrid doesn't touch, but any other priority either takes all the GPU or almost none of it. Cancers and other diseases are my moral priority, but getting the GPU as close to 100% is why I spent the money on the cards making it the financial priority for the moment. And the extra stats on the other side where I've spent almost 10 years doesn't hurt either. Maybe if I can think about it one day, I can figure out a way to make BOINC and Dnetc play nicer with each other.

---

1 Corinthians 9:16 "For though I preach the gospel, I have nothing to glory of: for necessity is laid upon me; yea, woe is unto me, if I preach not the gospel!"
Ephesians 6:18-20, please ;-)
http://tbc-pa.org

OK, so I see that the distributed.net GPU client can set its GPU runtime priority. I can't find any information out there on how to set the GPU priority in Windows like you can with the CPU. Does BOINC or the GPUGrid have any settings that can change the GPU runtime priority of it? Since I find that I need to run the DNetc at level 2, running BOINC GPUGrid at level 3 would then allow for BOINC to use as much of the GPU as it can force, then the DNetc will take the remaining percentage. If BOINC/GPUGrid does not have any way of changing or forcing this, does anyone know how to change it per task at the Windows OS level.

When I want to change the CPU runtime, I use Process Explorer. It is like Task Manager on steroids. It allows manipulation at program level, OS level, and CPU level and gives usage graphs and statistics. Now this also comes with "System Information" that gives graphs of your usages. Drilling into them, you can view your GPU engines. But even this program that can change all things program and CPU/OS level, it seems to not be able to manipulate the GPU in the same way (yet maybe.) Anyone? Anything? Even speculation if there is no clear answer.

---

1 Corinthians 9:16 "For though I preach the gospel, I have nothing to glory of: for necessity is laid upon me; yea, woe is unto me, if I preach not the gospel!"
Ephesians 6:18-20, please ;-)
http://tbc-pa.org