The speed is amazing.Going from 3800s to 2200s for single task,and 3500s for 2 task.That is on PCIe x16 v1.1 not on v3.0

Great work

Yesterday I upgraded my GPU to a factory over clocked GTX 660Ti.

I've run about 15 units, two at a time, and observed a speedup of about 700s compared to the BRP4 v1.25 on Einstein. At the same time the CPU-time per result has decreased by about 400s, thats more than half on my system.

v1.28 on Albert, x2, run time ~2170s and CPU time ~360s.
v1.25 on Einstein, x2, run time ~2900 and CPU time ~770s.

GPU-load also increased from ~80% to 95+%.

Great work!

---

How bout a single work unit?I want to see if a 660Ti is faster than a 580gtx/

How bout a single work unit?I want to see if a 660Ti is faster than a 580gtx/

I'd like to test that but as of right now I can't get any more work for BRP4, probably because the server status pages says 0 tasks to send...

---

BRP3Cuda32
GTX550Ti 1793 / 303 single wu
i3 win7/64 7.0.31 x16 slot

---

I ran the new CUDA 1.28 app via one of my Windows systems today. I have not been able to get much work today but the two tasks that ran via my GTX 580, completed at 834 seconds each. This is with one task running at a time. GPU load was at approximately 90-91% while running one task.

If memory serves me right, the previous application ran at around 1360 seconds per task with the 1.25 app via this system. This is a very decent improvement in performance. Thanks for the work put into optimizing the BRP4 applications.

Thanks for the feedback. That's actually a bit more of a speedup than I had expected based on some tests on slower hardware. Definitely in relative terms, the speedup is more pronounced on faster cards.

I will now install the Linux CUDA app on Albert, stand by for more tests. I'm eager to see those GTX 680 .... ;-)

Cheers
HB

---

How bout a single work unit?I want to see if a 660Ti is faster than a 580gtx/

Got hold of 2 BRP4-tasks and ran them one at a time on my over clocked GTX660Ti (Core@1201.9 MHz).
Run time: 1183.69 and 1175.39 seconds for an average of 1179.54 s.
GPU load ~84%.
This on Win7 x64, PCI-E 3.0x16.

---

Here are some preliminary numbers for the GTX 680.

One task per GPU

System #1 - Single GPU

x16 3.0 - 721 seconds

System #2 - Multi GPU

x16 3.0 - 785 seconds
x8 3.0 - 901 seconds

Overall, the performance looks great so far. I want to do some more testing with multiple tasks running at once, different PCI-E configurations, and with the CPU dedicated for BRP4 GPU only. The above tests were done with ~50% CPU load from running other CPU tasks at the same time.

I have a possible finding--not even a little bit sure--but am posting in case others might spot such a thing.

I've got two different hosts with the same GPU, a GTX 460. Neither has had downclocking problems for some weeks, but I found both downclocked severely today, with the problem persisting through system reboot.

It might just barely be possible that running the current Albert BRP1.28 CUDA ap, or on an even less likely note the Albert 0.29 Gamma Ray Pulsar application--or switching back and forth from those to the current Einstein applications was involved.

More likely something else in my system's history was the problem, but I thought I'd post the suspicion in case someone else sees something.

I'm not even sure what the true downclock frequency was, as different sources reported different numbers, but it was either 405 MHz or less--the reduction in power consumption and GPU temperature, while reporting exceedingly high GPU utilization, but making very slow progress on the WU was persuasive that downclocking was at hand.

---

I have sometimes the same issue.It only happens when i overclock the gpu +130Mhz from baseline.After several hours the screen goes blank and with gpu-z it shows core speed of 400Mhz then i proceed with reboot and everything goes to normal.
I think it some fail safe method that Nvidia uses.Also I noticed when boinc is not running the core speed drops to 50mhz and goes to 800 instantaneous if a demanding gpu work is needed

Another idea: What I'm always a bit suspicious about (BOINC-performance-wise) is OSX's power saving feature to automatically switch between a CPU build-in GPU (e.g. the Intel stuff in Sandy/Ivy Bridge CPUs) and the dedicated graphics card. Could this be the case that this feature is set differently on those two hosts perhaps?

Just to be clear, that feature can be turned off, so that the higher performance GPU is always used.

Also, be sure to be clear on which GPUs we are talking about. For example, there is the HD 5850 and there is the Mobility HD 5850. While the numbers are the same, they are completely different things. HD 5850 uses the Cypress PRO core, and is 2088 gflops. The Mobility HD 5850 is the Juniper chip, and is 800-1000 gflops, depending on the clock. The mobility versions are used in lap tops, including the Macbooks.

---

Dublin, California
Team: SETI.USA

After some time I decided to look for Albert again, this time with my nvidia. The reason was:
30-Aug-2012 08:53:39 [Albert@Home] Requesting new tasks for ATI
30-Aug-2012 08:53:44 [Albert@Home] Scheduler request completed: got 0 new tasks

To my surprise there was work:

30-Aug-2012 10:41:08 [Albert@Home] Started download of einsteinbinary_BRP4_1.28_i686-pc-linux-gnu__BRP3cuda32nv270

Hmmm. It looks like there is not only a 1.28 OpenCL binary and it looks like we can expect a newer CUDA binany. At the moment we have 1.24 in Einstein.

Bernd or Bikeman, can you please give a bit more information? :)
Thanks,
Stephan
PS: Ah ... I see that there was an bit of information. But mostly for Windows. Now let's see what my Linux box can do.

---

Hi!

Indeed, the 1.28 CUDA app will be released on Einstein@Home shortly ...probably today.

Cheers
HB

---

The older cards are also running well with the new version.

8800GT G92 512 MB - x16 slot @ 5.0 GT/s

1.28: 2940 seconds
1.24: ~3600 seconds