I have 2 480 GTX cards now completing beta work units and they seem to be very slow. I'm running W7 x64 and both cards clocked at 775mhz/1000mhz with fan on auto. These GPU clocks are stable after many hours of successful gaming. GPU usage on both cards is around 40%. I added the SWAN_SYNC=0 but so far each work unit is only using 3% of all of my processor (i7 920 with hyperthreading enabled, 3.4ghz overclock). My BOINC options allow for full CPU usage at all times.

I could really appreciate some help to get each work unit to use the entire CPU thread and not just 3%.

http://www.gpugrid.net/result.php?resultid=2259951

<core_client_version>6.10.50</core_client_version>
<![CDATA[
<stderr_txt>
# Using device 1
# There are 2 devices supporting CUDA
# Device 0: "GeForce GTX 480"
# Clock rate: 0.81 GHz
# Total amount of global memory: 1576468480 bytes
# Number of multiprocessors: 15
# Number of cores: 120
# Device 1: "GeForce GTX 480"
# Clock rate: 0.81 GHz
# Total amount of global memory: 1576468480 bytes
# Number of multiprocessors: 15
# Number of cores: 120
MDIO ERROR: cannot open file "restart.coor"
# Time per step (avg over 50000 steps): 24.510 ms
# Approximate elapsed time for entire WU: 1225.485 s
called boinc_finish

The task on GPU Grid don't use the CPU to do useful work, increasing this percentage will not substantially make the task run faster... mostly the CPU side is a "helper" thread to load data onto the GPU and remove results ... unlike SaH's ATI application or Einstein's CUDA application where substantial processing takes place on the CPU side ...

Danger30Q, your last task ran in 15min, which is faster than the 18min for the previous Betas. So you made some improvement whatever you did.

I have the same problem with my GTX470. I had to overclock it to get the time down from 1091sec (18min) to about 690sec (11.5min). But this is a bad fix!
MrS suggested my card was defaulting to some power saving mode and the clocks were not going back up correctly. Don’t know if the problem is to do with the drivers or the card, or what, but it is not the fault of Boinc, according to MrS.

Paul, the CPU usage can speed up the project. Hence a card on a board with a 1.6GHz AMD Athlon will not get through as many tasks as a card on a board with an i7-980X.
Also, Vista/W7 slows down the project compaired to XP and Linux is much faster as it appropriates a full core to support the GPUGrid task.

Does anyone know how to get each work unit to maximize a full thread on the CPU? I used the SWAN_SYNC=0 on Win7 x64 and my BOINC preferences are set to use 100% of CPU if needed. I'm still only at 3% CPU usage for each work unit. I think if I can get that to 12% (full thread usage with 8 total threads on the i7 920) then my work units should finish a bit faster.

I have the same problem with my GTX470. I had to overclock it to get the time down from 1091sec (18min) to about 690sec (11.5min). But this is a bad fix!
MrS suggested my card was defaulting to some power saving mode and the clocks were not going back up correctly. Don’t know if the problem is to do with the drivers or the card, or what, but it is not the fault of Boinc, according to MrS.

Paul, the CPU usage can speed up the project. Hence a card on a board with a 1.6GHz AMD Athlon will not get through as many tasks as a card on a board with an i7-980X.
Also, Vista/W7 slows down the project compaired to XP and Linux is much faster as it appropriates a full core to support the GPUGrid task.

Something that I ran across somewhere: this system started out doing beta WUs in ~1200 seconds. There's a setting in the NVIDIA control panel, under 3D Settings>Manage 3D settings, called "Power management mode". Setting this to "Prefer maximum performance" instead of the default "Adaptive" brought times down to below 1000 seconds. SWAN_SYNC=0 brought times down below 900 seconds, mostly around or below 850.

On the principle of "What's worth doing, is worth overdoing.", after seeing the improvement from SWAN_SYNC=0, I tried disabling hyperthreading. The one task that I ran with hyperthreading off did run faster, and with a shorter time per step, but not so much so that I was willing to continue the experiment.

[Question for project staff] Is SWAN_SYNC=0 just continuous polling?

tng, I just put the Fermi to have a go at your suggestion.

Boinc,
03/05/2010 02:15:24 NVIDIA GPU 0: (driver version 19621, CUDA version 3000, compute capability 2.0, 1280MB, 50 GFLOPS peak) :)

Installed driver,
03/05/2010 02:40:34 NVIDIA GPU 0: GeForce GTX 470 (driver version 19741, CUDA version 3000, compute capability 2.0, 1248MB, 726 GFLOPS peak)

Set option to max performance.

Ran a Fermi Beta,
Still taking about 11min.
3 May 2010 1:37:41 UTC 3 May 2010 1:51:13 UTC Completed and validated 656.59 649.57 187.28 280.92 ACEMD beta version v6.23 (cuda30)
Boinc still reading sleeping clock rates?
http://www.gpugrid.net/result.php?resultid=2262368

Used Asus SmartDoctor to up the engine to 707, then EVGA Precision to up the shaders and then MSI Afterburner to up the fan speed. Restarted Boinc,
03/05/2010 02:53:16 NVIDIA GPU 0: GeForce GTX 470 (driver version 19741, CUDA version 3000, compute capability 2.0, 1248MB, 1165 GFLOPS peak)

Started another task,
time down to just under 10min.

Nevermind, I got SWAN_SYNC now entered properly and each work unit now uses an entire thread all for itself. I also changed the Global Settings "Power management mode" to "Prefer maximum performance". That should help in crunching and in gaming.

tng, I just put the Fermi to have a go at your suggestion.

Boinc,
03/05/2010 02:15:24 NVIDIA GPU 0: (driver version 19621, CUDA version 3000, compute capability 2.0, 1280MB, 50 GFLOPS peak) :)

Installed driver,
03/05/2010 02:40:34 NVIDIA GPU 0: GeForce GTX 470 (driver version 19741, CUDA version 3000, compute capability 2.0, 1248MB, 726 GFLOPS peak)

Set option to max performance.

Ran a Fermi Beta,
Still taking about 11min.
3 May 2010 1:37:41 UTC 3 May 2010 1:51:13 UTC Completed and validated 656.59 649.57 187.28 280.92 ACEMD beta version v6.23 (cuda30)
Boinc still reading sleeping clock rates?
http://www.gpugrid.net/result.php?resultid=2262368

Clock rate of .81 GHz is standard for a GTX 470. I don't think BOINC reads the clock rate from the card, I think it just IDs the card and reports the clock rate based on that.

My first beta-test on gtx470 = OK
# 145 takes 6 min 19 secs
including systemfiles changed
boinc 06.10.50 says 0.17 cpu
taskmanager says 1.0 cpu
windows xp-pro

---

Ton (ftpd) Netherlands

[Question for project staff] Is SWAN_SYNC=0 just continuous polling?

yes


gdf

gtx480 now 5 min 21 secs
Time per step (avg over 50000 steps): 6.371 ms
# Approximate elapsed time for entire WU: 318.529 s

Better now?

---

Ton (ftpd) Netherlands

gtx480 now 5 min 21 secs
Time per step (avg over 50000 steps): 6.371 ms
# Approximate elapsed time for entire WU: 318.529 s

Better now?

Yes. In Linux running standalone it takes 5.8/5.9 ms. So it is almost there now.

gdf

My first beta-test on gtx470 = OK
# 145 takes 6 min 19 secs
including systemfiles changed
boinc 06.10.50 says 0.17 cpu
taskmanager says 1.0 cpu
windows xp-pro

Your first test on your GTX470 ran well, but the next two dropped to half speed.
My problem is that my card does not come out of half speed, ever. Basically it tries to default to power saving mode no matter what I do. The tools I have must be reading the maximum values, but the app never sees them.
I think the drivers are to blame.

PS. 0.81 is not normal - it is just typical of the wrong values being reported.

Clock rate: 1.30 GHz
This is what I clocked it to, but it is not performing like it should, so it must still be dropping to half that.

The first one was running "stand-alone".
All other jobs paused! 8 processor!

---

Ton (ftpd) Netherlands

It's like BOINC is only recognizing the core clock on the cards instead of the shader clock. All of my work units on my 480 GTX cards have used only 0.8ghz which is very close to 1/2 of my shader clock of 1550mhz. BOINC doesn't realize that the Fermi cards only have shader clock and no core clock.

For these present Fermi Betas,
a GTX480 should complete in about 5min 30sec,
a GTX470 should finish in about 6min 30sec.

My times are almost twice that, as are your times (Danger30Q).
We are both using Win7 64bit,
ftpd is using XP X86.
There might be something in that - perhaps the Win 7 drivers are not up to the job yet?

We both optimized as well as we could, so I'm going to pull some systems apart and setup a test XP system to try this GTX470 on. I hope there are still Betas around when I get it setup :)
- will let you know how I get on.

My problem is that my card does not come out of half speed, ever. Basically it tries to default to power saving mode no matter what I do. The tools I have must be reading the maximum values, but the app never sees them.
I think the drivers are to blame.

I seem to remember that you mentioned somewhere that your card was an ASUS. Have you tried uninstalling the ASUS programs that came with the card? They're the worst junk ever. I'm sure you've checked all the usual suspect settings like forcing max performance, yada yada. Another very likely possibility is that the ASUS BIOS is not behaving. I'd check with them for an update. Interestingly, I have 3 GT 240 GDDR5/512 cards that supposedly have identical specs, all different brands. The Gigabyte is just slightly faster than the MSI, and the ASUS is much slower than either. Switched them between machines and the ASUS is still always much slower. I've also had 2 ASUS 9600GSO cards. One was very slow and the other had a fan that failed after a few months. Also have an ASUS HD 4770 ATI card but that runs fine (have 8 HD 4770 cards from 4 different manufacturers and they all run perfectly). Maybe ASUS only designs well for GPUs that begin with the letter "A" :-)

3-5-2010 17:10:39 GPUGRID Sending scheduler request: To fetch work.
3-5-2010 17:10:39 GPUGRID Requesting new tasks for CPU and GPU
3-5-2010 17:10:43 GPUGRID work fetch suspended by user
3-5-2010 17:10:44 GPUGRID Scheduler request completed: got 0 new tasks
3-5-2010 17:10:44 GPUGRID Message from server: No work sent
3-5-2010 17:11:01 GPUGRID work fetch resumed by user
3-5-2010 17:11:20 GPUGRID Sending scheduler request: To fetch work.
3-5-2010 17:11:20 GPUGRID Requesting new tasks for CPU and GPU
3-5-2010 17:11:25 GPUGRID Scheduler request completed: got 0 new tasks
3-5-2010 17:11:25 GPUGRID Message from server: No work sent
3-5-2010 17:12:00 GPUGRID Sending scheduler request: To fetch work.
3-5-2010 17:12:00 GPUGRID Requesting new tasks for GPU
3-5-2010 17:12:05 GPUGRID Scheduler request completed: got 0 new tasks
3-5-2010 17:12:05 GPUGRID Message from server: No work sent
3-5-2010 17:12:40 GPUGRID Sending scheduler request: To fetch work.
3-5-2010 17:12:40 GPUGRID Requesting new tasks for CPU
3-5-2010 17:12:45 GPUGRID Scheduler request completed: got 0 new tasks
3-5-2010 17:12:45 GPUGRID Message from server: No work sent
3-5-2010 17:12:45 GPUGRID Message from server: Fermi-class GPU needed
3-5-2010 17:15:21 GPUGRID Sending scheduler request: To fetch work.
3-5-2010 17:15:21 GPUGRID Requesting new tasks for GPU
3-5-2010 17:15:26 GPUGRID Scheduler request completed: got 0 new tasks
3-5-2010 17:15:26 GPUGRID Message from server: No work sent

No more work for GTX260 on my HOME computer which may crunch all applications?

---

Ton (ftpd) Netherlands

Double check you home profile!
It looks like your GTX260 is using task your profile for Fermi's

Kev,

(Of course) i did that already. This morning it was working OK, but this afternoon .....

Perhaps beta's ask for fermi-gpu?

---

Ton (ftpd) Netherlands