Could someone tell me how much VRAM an NVIDIA GPU WU for Astropulse v.6 requires? I'd like to adjust the app_config.xml's in my machines to run multiple GPU WU's simultaneously based on the installed cards.

Thanks for any help you can provide.

Regards,
Steve

Could someone tell me how much VRAM an NVIDIA GPU WU for Astropulse v.6 requires? I'd like to adjust the app_config.xml's in my machines to run multiple GPU WU's simultaneously based on the installed cards.

Thanks for any help you can provide.

Regards,
Steve

Only one note : A GPU with 256MB VRAM would be limited to run one AP task at a time.

Guess your GTX 460 and your GT 550TI should be able to crunch 2 at a time.
The GT 430 eventually is too weak for two at a time. Maybe you need just to experiment on this.

In any case keep at least one CPU core free to feed the GPU(s).

In your ".../projects/setiathome.berkeley.edu/" folder there should be a "AstroPulse_OpenCL_NV_Readme.txt" file with further hints, e.g. available commandline options.

---

_\|/_
U r s

Actually, given that the CPU time for Nvidia AP GPU WUs is almost the same as the elapsed time, you really need 1 CPU per executing AP WU.

---

Roughly 530MB for two instances of AP on the one card. My monitor is driven by one card (GTX580) and uses circa 650MB with 2 instances of AP running, the other card (also a GTX580) only crunches and uses circa 530MB for the 2 instances of AP that are running on that card.

Hope that helps.

Lionel

---

diretly depends on setting app uses.
Memory usage varies.

---

SETI apps news
We're not gonna fight them. We're gonna transcend them.

First, I wanted to thank everyone for their responses.

My systems are all Linux/Nvidia and when I started getting my first Astropulse GPU WUs the other day, I noticed that they said something like (0.138CPU and 1 NVIDIA GPU). Since at that point I did not have an app_config.xml, I took that to mean the "native" preference for those WUs. Therefore, when I created my initial app_config.xml, I went with .2 for the CPU and .5 for the GPU as something that should be safe. That would give each WU at least 375 MB of VRAM, even on the smallest Fermi.

I checked my account this morning and so far all of the Astropulse WUs that have been finalized have all been accepted; no rejects.

It also struck me that the neighborhood of 300MB of VRAM seems to be something of a rule of thumb for any GPU WU; Seti, Einstein, or MW. Just an impression.

Anyway, there you have it.

Thanks again to everyone for their suggestions.

Regards,
Steve

Even though the common census has been to reserve 1 full core when running multiple AP WUs on GPUs, I've found that I can run with .5 CPU for each WU on my i7/950 w/ 2 x EVGA GTX660SC @ 2Gb; leaving me with 6 CPU cores for other processing. I have not noticed any difference in the time it takes to process a WU when upgrading from v6 -> v7. The max Vram that I seen is about 576Mb on the GPUs. BTW, you will probably find that the greater majority of people will disagree with me with the CPU reservations.

---

I don't buy computers, I build them!!

Even though the common census has been to reserve 1 full core when running multiple AP WUs on GPUs, I've found that I can run with .5 CPU for each WU on my i7/950 w/ 2 x EVGA GTX660SC @ 2Gb; leaving me with 6 CPU cores for other processing. I have not noticed any difference in the time it takes to process a WU when upgrading from v6 -> v7. The max Vram that I seen is about 576Mb on the GPUs. BTW, you will probably find that the greater majority of people will disagree with me with the CPU reservations.

That's probably because when most people look at your results they see the task using the same amount of CPU as Run time, All AstroPulse v6 tasks for computer 5501972. To most people that means it's using a full core no matter what you have assigned it. There are computers that don't use that amount of CPU time even though they are using the same driver as other people. I have tried the same driver and mine use ~100% unlike, say, Marks', All AstroPulse v6 tasks for computer 2645052. So, when you say yours isn't using a full core, but the results say otherwise, people would tend to disagree...

I wish I knew why some peoples nVidia cards didn't use 100% while other peoples do. You would think if you used the same driver with the same type card, the results would be the same. If you look around you will see that isn't the case, some use 100%, others don't. Yours appear to be using 100%.

FWIW, on the machine with the Fermi 550TI (2GB of VRAM), I have the app_config.xml set to .2 cpu and .2 GPU and I haven't choked any WU's so far.

Having said that, I haven't tried to do any sums to see if this is the most efficient setting to process the greatest number of Astropulse WUs in a given time. However, since I'm also not getting a lot of them.....

Regards,
Steve

Even though the common census has been to reserve 1 full core when running multiple AP WUs on GPUs, I've found that I can run with .5 CPU for each WU on my i7/950 w/ 2 x EVGA GTX660SC @ 2Gb; leaving me with 6 CPU cores for other processing. I have not noticed any difference in the time it takes to process a WU when upgrading from v6 -> v7. The max Vram that I seen is about 576Mb on the GPUs. BTW, you will probably find that the greater majority of people will disagree with me with the CPU reservations.

That's probably because when most people look at your results they see the task using the same amount of CPU as Run time, All AstroPulse v6 tasks for computer 5501972. To most people that means it's using a full core no matter what you have assigned it. There are computers that don't use that amount of CPU time even though they are using the same driver as other people. I have tried the same driver and mine use ~100% unlike, say, Marks', All AstroPulse v6 tasks for computer 2645052. So, when you say yours isn't using a full core, but the results say otherwise, people would tend to disagree...

I wish I knew why some peoples nVidia cards didn't use 100% while other peoples do. You would think if you used the same driver with the same type card, the results would be the same. If you look around you will see that isn't the case, some use 100%, others don't. Yours appear to be using 100%.

When I use .5 for both the CPU setting & .5 for the GPU count @ 80% of multi-processors (6 cores), there are 2 cores reserved for GPU processing of 4 WUs & 4 cores for CPU. If I set the CPU setting to 1, then the number of CPU cores drops to 2 for CPU processing, while the number of cores for GPU processing increases to 4. The time to process the GPU WUs remains approx. the same regardless of how many CPU cores I reserve. Since I see no apparent difference in processing time, I must conclude that what I am seeing in the BONC Manager is correct. The only time I see a noticable increase in the est. remaining times is when I take the machine out of hyper-threading mode, in which case those times nearly double. I also have to increase the CPU percentage to 100% to utilize both GPUs. Anything less will only allow one to process.

As I understand it there has not been an improvement in the processing of such tasks with the v7 implementation and as I recall the running times are consistent to those prior to the v7 implementation. I really don't care how long it takes to get my work done as long as the greater majority (99+%) are valid and the invalids are not caused directly by me. As to not running at 100%, I find that causes video lag in some of my applications and 80% is the sweet spot.

---

I don't buy computers, I build them!!

FWIW, on the machine with the Fermi 550TI (2GB of VRAM), I have the app_config.xml set to .2 cpu and .2 GPU and I haven't choked any WU's so far.

Having said that, I haven't tried to do any sums to see if this is the most efficient setting to process the greatest number of Astropulse WUs in a given time. However, since I'm also not getting a lot of them.....

Regards,
Steve

That settings would mean doing 5 wus at the same time.

Do you know a tool on linux to measure your GPU-load ?

---

_\|/_
U r s

FWIW, on the machine with the Fermi 550TI (2GB of VRAM), I have the app_config.xml set to .2 cpu and .2 GPU and I haven't choked any WU's so far.

Having said that, I haven't tried to do any sums to see if this is the most efficient setting to process the greatest number of Astropulse WUs in a given time. However, since I'm also not getting a lot of them.....

Regards,
Steve

That settings would mean doing 5 wus at the same time.

Do you know a tool on linux to measure your GPU-load ?

Urs,

I just checked the repository but couldn't identify anything. What I'm using right now to guide me are the GPU and CPU temperature monitor app that I installed on the toolbar and then the Nvidia x-config app.

On my regular non-Boinc machine with a Fermi 460, the temp is usually 38oC. On the Boinc machines (all cards are Fermi 430/460 or 550TI) that are running with .05CPU/.5GPU settings the temp run 64oC for the 430, 68oC for the 460, and 70oC for the 550TI. Although now that I think about it, I do have one of the BOINC machines with a 460 is routinely runs at 58oC though I can't explain why.

All machines have extra cooling in the form of side fans and CoolerMaster heatsink/fan combos. I basically guide on the temps. If I see them drop below those levels, I know that I've run out of GPU WUs.

Also the Nvidia x-config app on all of the machines is set to "adaptive" for the various speed settings.

I should add that my Boinc machines are strictly Boinc.

Regards,
Steve

Even though the common census has been to reserve 1 full core when running multiple AP WUs on GPUs, I've found that I can run with .5 CPU for each WU on my i7/950 w/ 2 x EVGA GTX660SC @ 2Gb; leaving me with 6 CPU cores for other processing. I have not noticed any difference in the time it takes to process a WU when upgrading from v6 -> v7. The max Vram that I seen is about 576Mb on the GPUs. BTW, you will probably find that the greater majority of people will disagree with me with the CPU reservations.

Your vid card usage is odd to me. How many WUs are you doing simultaneously?
My GTX 460s when running 2 x AP6 are running at about 288 and 190MB, repectively (1st is also monitor).

---

Even though the common census has been to reserve 1 full core when running multiple AP WUs on GPUs, I've found that I can run with .5 CPU for each WU on my i7/950 w/ 2 x EVGA GTX660SC @ 2Gb; leaving me with 6 CPU cores for other processing. I have not noticed any difference in the time it takes to process a WU when upgrading from v6 -> v7. The max Vram that I seen is about 576Mb on the GPUs. BTW, you will probably find that the greater majority of people will disagree with me with the CPU reservations.

Your vid card usage is odd to me. How many WUs are you doing simultaneously?
My GTX 460s when running 2 x AP6 are running at about 288 and 190MB, repectively (1st is also monitor).

I'm running 2 AP6 WUs each and at this post #1 is about 560MB & #2 about 430Mb #2 is attached to the monitor. I could go to 3 each and the time will extend by 30 - 40 min and consuming a max of 727 & 661 Mb respectively, but I've haven't tested that setting to any great degree, just a snap shot.

---

I don't buy computers, I build them!!

That seems bizarre - you are using > 2x as much VRAM as I for the same number of WUs (2). And the GPU running the monitor is using LESS than the other???

What OS are you running (I'm Win7x64)? And what are you using to measure the VRAM usage (I'm using EVGA Precision)?

---

That seems bizarre - you are using > 2x as much VRAM as I for the same number of WUs (2). And the GPU running the monitor is using LESS than the other???

What OS are you running (I'm Win7x64)? And what are you using to measure the VRAM usage (I'm using EVGA Precision)?

You are using standard params.

Cliff is using optimised params.
Unroll increases memory usage a lot.

---

With each crime and every kindness we birth our future.

That seems bizarre - you are using > 2x as much VRAM as I for the same number of WUs (2). And the GPU running the monitor is using LESS than the other???

What OS are you running (I'm Win7x64)? And what are you using to measure the VRAM usage (I'm using EVGA Precision)?

You are using standard params.

Cliff is using optimised params.
Unroll increases memory usage a lot.

My bad, the GPU connected to the monitor is #1, or in BOINC parlance -- device 0.

---

I don't buy computers, I build them!!