The long Fermi Work Unit completed in 22884sec (6h 21min), on XP x86, AMD 5200+, with no optimization, and running no CPU tasks:
2298270 1450742 8 May 2010 12:32:12 UTC 8 May 2010 18:58:06 UTC Completed and validated 22,884.27 7,483.39 7,954.42 11,931.63 Full-atom molecular dynamics v6.73 (cuda30)

So without using the swan_sync=0 variable, and using a lesser system (CPU) than before this task took about 32% longer.
- So a faster CPU and using the swan_sync=0 variable allows a 6.73 WU to finish 32% faster
(slightly better than the 26% observed for the 6.23 Beta)

- Now running another long WU (at stock) with the swan_sync=0 variable in place .


PS. Danger30Q,
When I had my GTX470 on W7 x64, and the same CPU as you, its performance was poor - similar to the performance you are seeing now; your quickest WU is 21,806. I am still seeing this when using no optimizations, but under XP x86 and using a lesser system (especially the CPU).

After trying everything I could think of with W7, I installed XP x86 on the same system (just different HDD), I set swan_sync=0 and found that my GTX470 completed a work unit much faster; 25% less time than your GTX480.

When I overclocked the card to 715MHz (with linked Shader rates) it completed a task (on XP x86, with swan_sync=0 on),
in 41% less time than your GTX480 on W7 64bit.
By my calculations a GTX480 should be 21% faster than a GTX470, so that is some difference!

I cant tell what setup you use, but I would guess you are crunching on your CPU and perhaps you are not using the swan_sync=0 variable?
- Think about this - A GT240 (£50 to £80) can do as much work as an Intel i7-980X (about £855).

I would strongly suggest that for now, and where possible, anyone with a Fermi should use XP or Linux, rather than Vista or W7, should use swan_sync=0 (XP), and should leave at least one CPU thread free to allow the GPU to excel.

Tomorrow we will have a fair idea how much the CPU matters when it comes to Fermi (swan_sync=0 now in place).

Provisionally (25% complete), it looks like about 20500sec to finish on the lesser system (and CPU) with swan_sync=0 enabled.
- This indicates the importance of the overall system performance, and in particular the CPU (just 17306sec on an 8 threaded system). So it would appear that the AMD 5200+ takes 18% more time than an i7-920, using the same GPU and operating system.
Althought this is not exact it is clear that the CPU makes a big difference!
It also strongly suggests that swan_sync=0 is an important factor (about 8%).

Good luck,

I do have SWAN_SYNC=0 enabled and both cards definitely have their own CPU thread. I'm not using my CPU for anything else except browsing the forums with Firefox. My CPU is only at 3.4ghz. My gpu usage on both cards is hovering around 55-57% at all times. Win7 x64 is just slow.

Is there anyway I can run my GTX 295 along with my (2) GTX 480s on an EVGA Classified motherboard? I remember hearing issues with the Classified not being able to run 4 gpus in GPUgrid but I hope I'm wrong.

The 480s replaced my GTX 295 and I may either sell it or put it in a separate Linux-based home server and run GPUgrid on it. Until I decide what to do with it, I'd love to run it along with my 480s. I have a Corsair 1000W power supply.

When I put the GTX470 into an AMD 5200+ system (without using swan_sync=0), it took 22,884sec to complete a task (11931 credits) which would earn 45K per day.
With swan_sync=0 in use, it took 20,325sec to complete a similar task. At that rate it would earn about 50.7K per day.
So, swan_sync=0 enabled on an AMD 5200+ system allows the task to complete 12.7% faster.

2298987 1451188 9 May 2010 0:52:49 UTC 9 May 2010 18:48:17 UTC Completed and validated 20,301.06 20,075.09 7,954.42 11,931.63 Full-atom molecular dynamics v6.73 (cuda30)
2298907 1451132 8 May 2010 12:36:05 UTC 9 May 2010 0:46:56 UTC Completed and validated 20,354.42 20,161.84 7,954.42 11,931.63 Full-atom molecular dynamics v6.73 (cuda30)

On an i7 system with similar setup (XP x86, swan_sync=0, native clocks, report tasks immediately) a task completed in 17,345sec (59K per day).
So, the faster system (i7) increased the speed significantly (by about 17%) compaired to a slower system (5200+).

Conclusions,
Dont use Vista or Win7 with a Fermi.
Put your Fermi into a fast system with a fast CPU.
Use swan_sync=0, and free up a core/thread (set Boinc to use 99% of the cores) so the GPU can benefit from the extra speed that will provide.

Now running a task at 715MHz GPU, 1430MHz Shaders, 1700MHz RAM:
The case it is in has better cooling, so at 66% fan the temp is 73 deg C.
The GPU % usage droped from about 64% to 60% when I upped the clocks!
Memory usage is only 382MB.
Task due to fin in 18,900sec; about 7% faster. On an i7 the same increase in clock rate sped the task up by 12%.

There is now a large (~1000) heap of production WUs on the acemd queue.

There is now a large (~1000) heap of production WUs on the acemd queue.

Except that now it seems to only send to the Fermi v6.73 app and won't send to the v6.72 app for the rest of us? Why has the much faster v6.72 app been disabled and the slow v6.03 app left active? Can't be to make Fermi look faster than it really is I hope :-(

There is now a large (~1000) heap of production WUs on the acemd queue.

Except that now it seems to only send to the Fermi v6.73 app and won't send to the v6.72 app for the rest of us?

Thanks much for turning v6.72 back on :-)

skgiven,

I wonder if I have to put swan_sync=0 for linux running machine?

---

You do not need to make any such environmental variable changes with the present Linux application. It already uses a full core by default. You can see this by looking at your GPU and CPU times to complete any 6.04 WU. Your daily return speaks for itself:
41,755 credits per day on a GTX275.

Are the WU for fermi-cards smaller?

Crunching one WU takes 2 hrs 16 min now (gtx480).

Correct?

Please allow to submit more WU then instead of 2 per card/cpu.

---

Ton (ftpd) Netherlands

These Shorter Fermi WU's bring more points.

2312870 1453463 11 May 2010 2:32:38 UTC 11 May 2010 7:28:13 UTC Completed and validated 8,796.39 8,486.95 4,503.74 6,755.61 Full-atom molecular dynamics v6.73 (cuda30)
2311711 1453000 10 May 2010 21:13:38 UTC 11 May 2010 5:01:24 UTC Completed and validated 8,710.02 8,430.89 4,503.74 6,755.61 Full-atom molecular dynamics v6.73 (cuda30)

D273r71-TONI_HERGunb1-0-100-RND0189_1
D309r257-TONI_HERGunb1-0-100-RND7525_1

The last few longer Fermi tasks (IBUCH_0511_pYEEI_long) would only allow me to get about 53.8K per day (on this slow but OC'd system) and only used about 60% GPU.

These new shorter tasks would allow me to get 66.7K per day (on the same slow system).

The same card with the same clocks in an i7 sould get 78K per day.

Has there been a Fermi app improvement or is it just that these TONI_HERGunb tasks run particularly fast (GPU usage is now about 82%)?
We are talking about a 24% improvement.

skgiven,
thanx man :-) I thought there is way (rather then get fermi) to improve my RAC.

I'm keeping eye on this thread to know when I should get fermi :-)

---

I'm not sure if it's the newer work units but my GPU usage on both of the 480s is around 70%. A few days ago, I was lucky to hit 60%. Excellent job if this is the result of Fermi work unit improvements.

Performance for a reference molecular system (DHFR):
GTX275 9.0 ms/step (cuda3.0)
GTX275 8.5 ms/step (cuda2.2)
GTX470 7.1 ms/step (cuda3.0)
GTX480 5.9 ms/step (cuda3.0)

Running temperature of about 92 degrees for all of them.

gdf

In the latest code we reached:
GTX480 5.45 ms/step (cuda3.0)

gdf

Excellent improvements, we look forward to more :cheers:
Those stats are for stock shader settings and stock voltages?
I'm also assuming the are from Linux systems? Can you share with us which flavor and version of Linux and CPU/ RAM specs.

---

Thanks - Steve

Excellent improvements, we look forward to more :cheers:
Those stats are for stock shader settings and stock voltages?
I'm also assuming the are from Linux systems? Can you share with us which flavor and version of Linux and CPU/ RAM specs.

Fedora core 10, GTX480 at stock clocks, X58 chipset and i7 2.66Ghz.

gdf

So then is it possible for us to run Fermi's on Linux now? I didn't see an annoucement.

So then is it possible for us to run Fermi's on Linux now? I didn't see an annoucement.

wake up man :-) I'm crunching GPUGRID on ubuntu a year already :-)

---

He was referring to the Fermi capable app being released to the Linux public.

MrS

---

Scanning for our furry friends since Jan 2002

ouch... sorry

---