Yes I have as I have bought a refurbished rig with 2 Xeon's no HT (not possible), and when running a GPU task and none CPU, GPU use gets a steady load. When adding a core at a time for CPU crunching, the GPU load drops, to below 35% with 7 cores and to zero with 8 cores crunching CPU. Then only very limited a core give some time to the GPU WU.

Okay, I understand now, I didn't realize that the new rig you bought didn't have Hyperthreading. I think it's strange that only the Intel processors experience this throttling and the AMD processors don't (at least in my case). I have removed 2 cores from BOINC so they can only be used by the video cards and nothing else in BOINC and I'm assuming everyone else has done this too.

Intel i3/5/7 CPUs can reach VERY high temps (~90C) quite easily with their stock Intel coolers! I don't understand why 70C under continuous crunching load is considered too hot or toasty! Isn't a difference of 20C from the consciously Intel-selected upper operating temperature, a safe distance?

As evidence to back this up I have my home server, crunching on WCG for years now, day in, day out, at temps ranging from ~55 to ~70.

I think people tend to over-react or go the extra mile(s) on "optimal" cooling without a real reason.

---

Intel i3/5/7 CPUs can reach VERY high temps (~90C) quite easily with their stock Intel coolers! I don't understand why 70C under continuous crunching load is considered too hot or toasty! Isn't a difference of 20C from the consciously Intel-selected upper operating temperature, a safe distance?

As evidence to back this up I have my home server, crunching on WCG for years now, day in, day out, at temps ranging from ~55 to ~70.

I think people tend to over-react or go the extra mile(s) on "optimal" cooling without a real reason.

Your are quite right about the temperatures, but with liquid cooling 71°C is indeed toasty.
Secondly running on such high temperatures will reduce the lifespan of a CPU significantly and more over circuit boards. So good/efficient cooling is a must for crunchers, especially 24/7 ones.

---

Greetings from TJ

TJ & skgiven,

OK, I am starting to catch on thanks to you guys help. I am quite certain that in compute preferences I told each project that they could use 100% of the processors 100% of the time. From what you guys have said I can go in there and set Rosie to 7 cpu cores and put GPUG on 1 core. That would certainly allow GPUG to go to its full potential. Right now when I click "run based on preferences" I see 8 Rosie tasks running and 1 GPUGRID task running. The status line for the GPUG task says "Running (0.721 CPUs + 1 NVIDIA GPU". If I change my preferences and lower the percentage given to Rosetta the GPU should load up better. Good to know. On the other hand I'm quite happy to run things the way that they are. I like keeping the GPU temp down as far as possible.

I have set under Tools, Computing preference, Processor usage to 75%. This results in one core doing nothing. In my case it results in slightly higher GPU load. This setting works for all tasks that run on the CPU not just Rosie. As far as I know you can not set that per project, you can only give a project more time by setting its share higher. But this will not affect CPU usage.
However you have a cool case, with cool I mean cool like the Americans use this word often with something is great :)

Secondly in EVGA Precision you have the option "GPU clock offset" this can be plus or minus. You can also click on "voltage" on the left site of the program under "test" and "monitoring". A new window will pop-up and you can lower the voltage. EVGA has neat software, that is one reason I like that brand.

I don't get it. Are you sure about lowering the voltage?? When I click on Voltage a new window pops up but lowering the V is not possible. All you can do is click overvoltage and then drag the arrow up to raise the voltage. My software revision is the new one: 4.2.0.

Yes same software here on two rigs. When I click on Voltage I get the same pop-up window and can draw the arrow with the mouse up and down from 1150 to 825 mV.
I can set i.e. at 900 mV, click on apply and then click on the red cross and it works. But I have other cards, that could have to do with it. Nothing higher than a GTX660.

---

Greetings from TJ

TJ is correct, just because the CPUs can hit and run at 70, doesn't mean they should 24/7/365. It WILL shorten the lifespan.

TJ-

Thanks for that info on voltage in the EVGA software. I have tried and tried, but can't get the arrow to go down. I have dragged on the arrow and dragged on the bars, nothing seems to work. There is a GPU tweaker that came with my Asus motherboard. I noticed that it won't allow me to lower the voltage either. I think this might be because I have the super-clocked edition of the card. I'm not going to worry about it- the card is working well and running cool anyway.

The 320.x driver is probably preventing this, going by the 60-odd page manual.

---

FAQ's

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

You need to remember that the GTX770 has a TDP of 230 watts while a GTX680 has a TDP of 195 watts, they upped the default max voltage from 1.175v to 1.187v.

Mines still on air at this point and averages 57° to 63° with the ACX cooler, I wish they would hurry up and release the next driver already.

Edit: SK, did you get a chance to read that manual? Probably some lite reading, eh?

Regarding the voltage: as you said under crunching load 1.187 V are displayed. This is actually the voltage for about the highest turbo bin. If you set a voltage of 1.175 V in Precision, this should limit the maximum used. But as you say, your card runs pretty well by now, so I'd rather use that headroom left in the chip for some OC without voltage increases. A good 50 MHz should surely be left. Or us the power target to drive power consumption down and efficiency up, if need. In tihs case the card will choose lower clocks & voltages itself.

BTW: I'm also glad EVGA makes the nice Precision software (actually a GUI for RivaTuner made by someone else, just as MSI Afterburner). And that it also works with my non-EVGA cards!

Regarding the temperature: higher temperatures do reduce a chips lifetime. It's a guaranteed, continous process. The rule of thumb is half the lifetime for every 10°C higher. And while it's true that manufacturers allow their CPUs and GPUs to hit ~90°C, this doesn't mean it's good this way. They're simply not assuming 24/7 load for consumer-grade hardware.

If this lifetime reduction will matter at all within the useful life of the hardware considered is a different question entirely and can not be answered generally. CPUs have historically been able to take quite a beating before one starts to notice the degradation (the chip will reach lower and lower clocks and similar conditions, or need more voltage to reach the same levels as before).

On the other hand I've burned out my first Radeon X1950Pro crunching Folding@Home within a few months of running 24/7 at ~70°C. There was surely some bad luck involved here, but I wouldn't want to run higher than 70°C under sustained load.

GPU-Grid performance vs. CPU threads: in my testing I reached full performance on my single GPU when I limited BOINC to 3 CPU WUs running on my Quad i7 with HT on. That means a full physical core dedicated to support the GPU. Yet I'm running at 100% CPU load now because I like to support Einstein and HT is especially efficient there, as well as my 22 nm CPU.

Note that on Kepler GPUs (pretty much all we're talking about now) GPU-Grid always uses a full CPU thread, despite that unfortunate "0.7xxx" value being shown by BOINC. This means by default BOINC will launch 8 CPU threads along GPU-Grid (on an 8-core machine) and overcommit the CPU. To fix this one can set it to "use at most 99% of CPUs", but then the CPU will not be fully used if GPU-Grid is down. Or one can place a file called "app_config.xml" into the GPU-Grid project folder with the following contents:

<app_config>
<app>
<name>acemdbeta</name>
<max_concurrent>9999</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
<app>
<name>acemdlong</name>
<max_concurrent>9999</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
<app>
<name>acemdshort</name>
<max_concurrent>9999</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
</app_config>

Won't make much of a difference, though.

MrS

---

Scanning for our furry friends since Jan 2002

TJ is correct, just because the CPUs can hit and run at 70, doesn't mean they should 24/7/365. It WILL shorten the lifespan.

I ignore this most of the year and run at about 68 to 78 degrees, but we are pushing 25 degrees Celsius in Hammerfest (Northernmost city in the world) so I've had to pause my cpu and gpu computing (normally I heat my house almost entirely on computing).
This leaves room for possibly upgrading without it affecting computing contribution. On that note, have you concluded anything profound in this thread? Or must we just wait and see for the first person to buy them all and pit them against each other?

Off-topic: Can AMD be used in GPUgrid yet? I have not stayed updated on the matter. except one thing, I have noticed some very cheap AMD cards trance my 560Ti in various OpenCL applications.

780 doesn't work yet. 770 is slightly faster than a 680 (but the jury is out on the performance/watt side).

Off-Topic Question - No

---

FAQ's

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

Hm, thanks. Then I'll have to wait a bit until 780 works well before I decide.

Hi MrS-

Thank-you for all of your input. I understand this:

Note that on Kepler GPUs (pretty much all we're talking about now) GPU-Grid always uses a full CPU thread, despite that unfortunate "0.7xxx" value being shown by BOINC.

But I do not understand this, and why it needs to be fixed:

This means by default BOINC will launch 8 CPU threads along GPU-Grid (on an 8-core machine) and overcommit the CPU.

I typically run Rosetta along with GPUG. I have noticed that if I suspend Rosetta and let GPUG run all by itself the GPU load goes up from 82% to 89-90%. A substantial amount. So I thought OK, I'll fix this and get GPUG going even faster.

To begin with I don't understand why Rosie is having an effect on GPUG because Rosie does not touch the GPU at all. I have verified this. (If I suspend GPUG and let Rosie continue the GPU load goes to zero). Apparently, as you were saying, there is a certain load that GPUG requires from the CPU, typically 1 core. So I go to Rosetta's page and adjust my compute preferences, making sure to update Rosie in BOINC, so BOINC knows what to do, but this has had no effect. Right now I have Rosetta's preferences set to use 1 processor for 25% of the time, but BOINC is still ramping up to 100% CPU usage on all cores when running Rosetta and GPUG together. My BOINC preferences are set to use 100% of the processors for 100% of the time, but I'm sure the way to turn down Rose is to do it on Rosie's page, right?

Does anyone have any suggestions?

Ronny-
Run your computers at 70 or 80 degrees and heat your house with the heat. That is amazing. Sounds to me like a study in efficiency. Well done!

Rosetta doesn't use the GPU, but GPUGrid does use the CPU.

Set Boinc locally to 99% and you should see a slight increase in GPU usage and WU performance. GPU usage would increase further if you set it to 75% or 50%, but most people are happy with 99% (7 CPU tasks on 7 CPU threads + 1 CPU thread to feed the GPU, basically).

---

FAQ's

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

Right now I have Rosetta's preferences set to use 1 processor for 25% of the time, but BOINC is still ramping up to 100% CPU usage on all cores when running Rosetta and GPUG together. My BOINC preferences are set to use 100% of the processors for 100% of the time, but I'm sure the way to turn down Rose is to do it on Rosie's page, right?

Does anyone have any suggestions?

Ronny-
Run your computers at 70 or 80 degrees and heat your house with the heat. That is amazing. Sounds to me like a study in efficiency. Well done!

It is not that you have to configure the participation of each project in the project preference page, this does only influence the priority of how many work is send from each project. In my case I have a priority project for CPU and GPU respectively: Top priority climateprediction.net 100%, secondary priority malariacontrol.net 1% so my CPU gets only work if either climateprediction.net does not have or the computer has to fulfill the ration of 100/1.
We do refer to “My BOINC preferences” in the BOINC Manager: So I do set “to use 99% of the processors for 100% of the time” as I do have one GPU in the system, so of my entire cores one core will be free to feed my GPU and all others are used in my CPU project. Or in your case you have to free 3 cores or threads to feed your 3 GPUs. This will make run GPUGRID project faster and your system in general will more stable.

Glad I asked about that! I reset my local preferences to compute on 99% of the processors and it reduced Rosetta to only 7 tasks running and one GPUG task running.

Hello: The GTX 780-TITAN GPUGRID work already, or is there any provision for this. Greetings.

Look over there for updates [there haven't been any, as the developer is away].

MrS

---

Scanning for our furry friends since Jan 2002

To fix this one can set it to "use at most 99% of CPUs", but then the CPU will not be fully used if GPU-Grid is down. Or one can place a file called "app_config.xml" into the GPU-Grid project folder with the following contents:

I have tried this on my "big system" (new PSU is on its way to fit two AMD GPU's) with Einstein@home and what I saw was that the GPU load went to 0 (zero) and then occasionally increases to 24% load. The WU in the BOINC taks list didn't also make progress. So I set CPU use back to 90% again to lease one core free.

---

Greetings from TJ

Mhh, something obviously went wrong there, but I don't know your configuration well enough to make any educated guesses. In my post I assumed a standard setup as starting point, that is 100% CPU use and 1 GPU. Maybe that doesn't apply to your system?

MrS

---

Scanning for our furry friends since Jan 2002