I just finished my first WU via the 580. Run time 9176 sec. I am not sure how good that is or not.

This is good. If you want maximum performance, you should dedicate one CPU core to feed the GPU. Use the SWAN_SYNC=0 environmental setting to achieve this. (control panel -> system -> advanced tab -> environment variables -> new (system variables) -> variable name: SWAN_SYNC value: 0 -> restart windows)

I just finished my first WU via the 580. Run time 9176 sec. I am not sure how good that is or not.

This is good. If you want maximum performance, you should dedicate one CPU core to feed the GPU. Use the SWAN_SYNC=0 environmental setting to achieve this. (control panel -> system -> advanced tab -> environment variables -> new (system variables) -> variable name: SWAN_SYNC value: 0 -> restart windows)

Thanks for the tip. I just added this environment variable.

This is the fastest result ever seen on GPUGRID 3.9 ms/step on the DHFR workunit:
http://www.gpugrid.net/result.php?resultid=3326415

However, it should be able to get to 3.3 ms/step (just guessing, as we don't have any GTX580 yet). Maybe, boosting the priority as suggested.

gdf

How about 3.747 ms / step?

I paste it here, before it'll be gone:

<core_client_version>6.10.58</core_client_version>
<![CDATA[
<stderr_txt>
# Using device 0
# There are 2 devices supporting CUDA
# Device 0: "GeForce GTX 580"
# Clock rate: 1.61 GHz
# Total amount of global memory: 1610153984 bytes
# Number of multiprocessors: 16
# Number of cores: 128
# Device 1: "GeForce GTX 480"
# Clock rate: 1.40 GHz
# Total amount of global memory: 1610285056 bytes
# Number of multiprocessors: 15
# Number of cores: 120
SWAN: Using synchronization method 0
MDIO ERROR: cannot open file "restart.coor"
# Time per step (avg over 2000000 steps): 3.747 ms
# Approximate elapsed time for entire WU: 7494.516 s
called boinc_finish

</stderr_txt>
]]>

I think it's very hard to reach 3.3 ms/step, because my GTX 580 runs at 95-96% GPU usage already. Maybe we can gain another 0.1-0.2 ms/step with heavy overclocking of the GPU (some bigger GPU cooler required to do this and maybe the new Core i3, i5, i7 processors, wich will be released in 2011/Q1).

Your 3.747 is the fastest result I have seen, and I did check a few systems.

You are probably correct in that if you had a faster system you might see slightly faster results and higher GPU utilization. To get right down to 3.3 (12% faster) you might also need an X64 system and to do no CPU work at all.
If you use GPUZ or one of the overclocking tools you could check to see if freeing up additional CPU cores makes any difference; if the GPU utilization rises you might want to run one task like that just to see what the optimum performance is on your system.

3.614 ms/step

It's happened overnight, so I couldn't watch the GPU usage.

<core_client_version>6.10.58</core_client_version>
<![CDATA[
<stderr_txt>
# Using device 0
# There are 2 devices supporting CUDA
# Device 0: "GeForce GTX 580"
# Clock rate: 1.70 GHz
# Total amount of global memory: 1610153984 bytes
# Number of multiprocessors: 16
# Number of cores: 128
# Device 1: "GeForce GTX 480"
# Clock rate: 1.40 GHz
# Total amount of global memory: 1610285056 bytes
# Number of multiprocessors: 15
# Number of cores: 120
SWAN: Using synchronization method 0
MDIO ERROR: cannot open file "restart.coor"
# Time per step (avg over 2000000 steps): 3.614 ms
# Approximate elapsed time for entire WU: 7227.063 s
called boinc_finish

</stderr_txt>
]]>

Excellent, you just shaved off another 0.133 ms/step, raised the bar a notch, and showed that the GTX580 can operate the most demanding tasks (GIANNI_DHFR1000) while overclocked by 10.7%. A very good card.

I've made a little calculation:
I've rised the GPU's clock 6.25% (1.7GHz/1.6GHz=1.0625)
The average time/step went down 3.55% (3.614ms/3.747ms=0.9645) in other words the processing is 3.68% faster (3.747ms/3.614ms=1.0368)
So these numbers are showing that other components really limiting the processing speed (as I supposed it before). If they weren't limiting it, the average time/step should be now 3.527ms (=1.6GHz/1.7GHz*3.747ms).
Taking the acceleration lost in account, to achieve 12% performance boost to reach 3.3ms/step, the GPU have to be overclocked 20.4% (6.25/3.68*12=20.38) (1.9264GHz shader and 963.2MHz core frequency) or more (because the other limiting factors could be higher at higher GPU clocks). I won't try this while I don't have some huge GPU cooler on my GTX 580. (In January I'll have)
But I have a faster CPU in my other PC, so I could swap the GPUs, or the CPUs, to find out how much help a faster CPU is. But to shut down both of my PCs at the same time for an hour (or longer), is a really hard thing to do.

Your 4% loss from 1h shutdown would be offset by a day or two's increased performance ;)

Presently my GTX470's are only OC'd by 8% and run the GIANNI_DHFR1000 tasks without failure (5.325ms per step). So your GTX580 is 47% faster. At 19% GPU OC these tasks failed on my GTX470, but the other tasks ran well. So if you can successfully run GIANNI_DHFR1000 tasks the rest should also run. They make for good stability test work units

PS. From reference shader freq. (1536MHz), your card is overclocked by 10.7%

Could well be that your GPU utilisation dropped a little upon further OC'ing. Or that GPU memory is holding you back. Generally I would have expected a much higher improvement going from 1.6 to 1.7 GHz.

MrS

---

Scanning for our furry friends since Jan 2002

I've removed my PC from it's case in order to make a new ventilator inlet at the bottom (the lower GPU cooler could not breath in enough air, because it's right next to the bottom of the case). The temps vent down of course, so I've raised the GTX580's clock to 900MHz. It has failed a couple of tasks, so I've raised the GPU's voltage by 13mV (to 1.063V). It's seems to be ok since then. It's completed a _KASHIF_HIVPR_n1_unbound_ task in 2 hours (7.203 ms/step). The max temp was 65°C at 95% GPU usage. (standard GTX580 cooler @78% plus a 12cm cooler placed right to the top of the two cards) I hope I'll receive a GIANNI_DHFR task soon.

I've set a new world record in crunching GIANNI_DHFRs: 3.441ms/step

<core_client_version>6.10.58</core_client_version>
<![CDATA[
<stderr_txt>
# Using device 0
# There are 2 devices supporting CUDA
# Device 0: "GeForce GTX 580"
# Clock rate: 1.80 GHz
# Total amount of global memory: 1610153984 bytes
# Number of multiprocessors: 16
# Number of cores: 128
# Device 1: "GeForce GTX 480"
# Clock rate: 1.60 GHz
# Total amount of global memory: 1610285056 bytes
# Number of multiprocessors: 15
# Number of cores: 120
SWAN: Using synchronization method 0
MDIO ERROR: cannot open file "restart.coor"
# Time per step (avg over 2000000 steps): 3.441 ms
# Approximate elapsed time for entire WU: 6882.250 s
called boinc_finish

</stderr_txt>
]]>

I think my CPU is also part of this success. I'm using a C2Q 9650 @ 4.25GHz (it's running at this speed at standard CPU voltage)
BOINC reports the Measured floating point speed 4438.87 million ops/sec
Some WUs failing on this GTX 580 @ 900MHz, so I think I will lower the GPU clock in the future.

"How fast do you want to get the wrong result?" is not the question we're trying to answer ;)
Anyway, very nice scores!

MrS

---

Scanning for our furry friends since Jan 2002

Records are made to be Broken at any Cost ... lol

Perhaps the 3.441ms/step record will fall when someone uses a Sandy Bridge CPU on Linux, with some light OC'ing.

EVGA has a 2-slot water cooled 580 out, street price in Akihabara, Tokyo is 70,000yen or over $700, a ¥20,000/$200 premium.

FYI, courtesy of google translation:

Water-cooled GeForce GTX card debut for models 580, EVGA's "GeForce GTX 580 FTW Hydro Copper 2 (015-P3-1589-KR)" was released.

　In addition to standard have been overclocked, it also features one slot is a card thickness. Street price is around 70,000 yen (see " new products this week found "for details.)

　This product is a liquid cooling system as standard GeForce GTX 580 Video Card. Are covered by a copper cooling head chrome plating across the board, PCI Express x16 connector on the other side of the fitting are also provided.

　Best of all, the point has been slimmed down one slot from the normal models for two-slot card thickness. Despite having to take a water cooling system is introduced, and likely choice for those who want to save expansion slot space.

　The main specifications, the installed memory GDDR5 1,536 MB, the core clock is 850MHz, shader 1,700 MHz, Memoridetareto 4,196 MHz (typically core 772MHz, shader 1,544 MHz, the memory 4,008 MHz), equipped with video terminals and Mini HDMI DVI ( group 2), card size is 111.15 × 266.7mm. The external power supply terminal is equipped with one group for each 6-pin and 8-pin type like normal models.


□ GeForce GTX 580 FTW Hydro Copper 2 (EVGA)

http://www.evga.com/products/moreinfo.asp?pn=015-P3-1589-KR

It looks really good, and should give good performance, but the connectors are vertically mounted, meaning that the tubing will protrude into the next slot. Pity they did not top/rear mount the connectors. pic

I just installed a new GTX 580 and ran a task to see how it performs.

Any idea why this task took so long? I have a second in progress now, and it looks like it will take the same amount of time. GPU load is ~98%. SWAN_SYNC=0. I have preferences set so that BOINC is using only 7 of 8 CPU cores. What am I doing wrong?

One thing that sticks out is the difference between the value for "Run time", and the value for "Approximate elapsed time for entire WU". What could cause those two values to be so different?

http://www.gpugrid.net/result.php?resultid=3522674

Run time 62752.374998 (17 hours)
CPU time 62631.21
SWAN: Using synchronization method 0
# Time per step (avg over 575000 steps): 21.622 ms
# Approximate elapsed time for entire WU: 27027.548 s (7.5 hours)

---

Reno, NV
Team: SETI.USA

Hi zombie67, that is FAR too slow for a GTX580. Something is very wrong!

Check your available RAM and HDD usage, just in case.

Right click on the desktop, click NVidia control panel, Manage 3D settings, Global Settings, Power management mode, Prefer Maximum Performance. Restart the system.

If you are running some very CPU intensive apps consider freeing another CPU thread; some CPU tasks want more CPU time than they can get, so in Boinc check the difference between CPU time and elapsed time, in case you are running some very hungry CPU tasks.

Did you install the driver after the card and then do a restart?

Hi zombie67, that is FAR too slow for a GTX580. Something is very wrong!

Check your available RAM and HDD usage, just in case.

Using only 1gb (of 8gb) of RAM. 70gb of 146gb free disk space.

Right click on the desktop, click NVidia control panel, Manage 3D settings, Global Settings, Power management mode, Prefer Maximum Performance. Restart the system.

I tried that, but it is not possible. On the Global settings tab, power management is not listed at one of the options. On the Program Settings tab, there are two problems: 1) you have to select the Program. BOINC/CUDA/GPUGRID/ACEMD2 none of these are listed. 2) Even if it was listed, the only option choice under Power Management is "Adaptive". In any case, according to GPU-Z, the GPU is not being throttled. It is running at full speed, and full load.

If you are running some very CPU intensive apps consider freeing another CPU thread; some CPU tasks want more CPU time than they can get, so in Boinc check the difference between CPU time and elapsed time, in case you are running some very hungry CPU tasks.

I actually ran most of that task using only 6 of 8 CPU cores. There was plenty of idle CPU cycles.

Did you install the driver after the card and then do a restart?

Yes. Before installing the card, I uninstalled the old card driver, then ran driver sweeper, then installed the latest driver from nVidia's site (263.09). After each step, I rebooted.

FWIW, on other CUDA projects, the card runs at equivalent speeds to other machines with the GTX 580. It is only this project where I am having this problem.

Edit: I think the differing values for CPU time are a clue. I just don't know enough about the app to understand what it means.

---

Reno, NV
Team: SETI.USA

I still think the card is throttling back.

The CPU time/ Run Time is consistant with the use of swan_sync=0 - this is fine.

Although your second task was faster 16ms per step, this is still wellshort of what it should be; a similar task on one of my GTX470's takes about 14ms per step at ref speeds, and I have seen 12ms per step for a GTX580 on Win7 (which is slower).