You didn't list x38g + 26x.xx drivers time. There were improvements from x32f so listed comparison is incorrect.

3rd AP only lasted a few secs but was validated http://setiathome.berkeley.edu/result.php?resultid=2001592219

4th AP http://setiathome.berkeley.edu/result.php?resultid=2001584832

Took quite a longer time than the 2 first, this time I also ran a MB on GPU and 4 CPU MBs, not sure if that was why it took longer.

Using a PhenomIIx4 9550@2.35ghz. GTX560ti@ 935-1870-2100 Driver 266.66

WU was doing .901% a tick at 4.79% blanking with gpu usage jumping between 8% to 90%.
I tried using different command line values to try and get a better %/tick but there was no increase or decrease except for Mem usage.

Using <cmdline>-instances_per_device 1 -hp -unroll 8 -ffa_block 6148 -ffa_block_fetch 2048</cmdline> but have tried up to
cmdline>-instances_per_device 1 -hp -unroll 16 -ffa_block 24576 -ffa_block_fetch 12288</cmdline> and a few other values between.

I can run 2 at a time but have to keep unroll at 8 so memory doesnt go too high.It is using 853 with 2WU about 490 with 1.Ruuning the 2 WU uses a little more of the gpu% but still jumps around between 15 and 95%.CPU uses 10 to 40%.

You didn't list x38g + 26x.xx drivers time. There were improvements from x32f so listed comparison is incorrect.

Cant give u that much data about it cause of the shorty storm or none WUs at all the last time. But this is the actual config of my Phenom (... which im actual run dry cause main reason this comp exists is not given at the moment, going to switch it off).

On the i7 this combination was used also to short to give good information (shortys/empty cache).

Sorry

greetings

Chris

---

- Performance is not a simple linear function of the number of CPUs you throw at the problem. -

The 560s have 8 compute units so i´d say unroll 4/8/12 are the values to try.

Thread_Block 6148 is wrong should be 6144 to be honest.

I would go with Thread_Block 8192 Thread_Block_Fetch 2048 and unroll 12.

---

With each crime and every kindness we birth our future.

...
WU was doing .901% a tick at 4.79% blanking with gpu usage jumping between 8% to 90%.
I tried using different command line values to try and get a better %/tick but there was no increase or decrease except for Mem usage.
...

The top level loop for AP processing runs 111 iterations, so you won't change that 1/111 = 0.9009009%/tick. Making it tick faster or slower is of course possible...

                                                                   Joe

Mike, if the 460's have 7 SM's, does an unroll of 7/14 make sense for these cards?
I've processed one WU with the 460 at the default of 10, but my wingman hasn't reported in yet. The output looks good as far as I can tell. The odd number of SM units doesn't make for easy integer values for unroll.

Cheers, Keith

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Finally completed my first couple, and one already validated. Good run times out of this little GTS450. Using 266.58, unroll 12, low priority, everything else default.

Task, Run Time, CPU Time, Blanking
2000647746 5,959.56 1,272.56 2.39
2000647748 5,735.61 1,059.23 0

The only problem is a little screen lag. Next I'm going to try Claggy's -ffa_block 6144 -ffa_block_fetch 1536. If that doesn't do it I may try -unroll 8 to see if it helps.

First correct validation against one stock client:

http://setiathome.berkeley.edu/workunit.php?wuid=785187633

---

- Performance is not a simple linear function of the number of CPUs you throw at the problem. -

Mike, if the 460's have 7 SM's, does an unroll of 7/14 make sense for these cards?
I've processed one WU with the 460 at the default of 10, but my wingman hasn't reported in yet. The output looks good as far as I can tell. The odd number of SM units doesn't make for easy integer values for unroll.

Cheers, Keith

Unroll of 7 definitely to low for 460.
450 can handle unroll 10 perfectly so i´d try between 12 and 14.
Doesn´t have to be integer.

My HD 5850 has 18 CU´s and unroll 12 is just perfect.

Note Thread_Block and Thread_Block_Fetch not shown in stderr.
Maybe you got a typo in appinfo.

For 450 i suggest 6144/1536
460/470/480/560/580 should use 8192/4096

If you get screen lags reduce Block_Fetch to 2048.

---

With each crime and every kindness we birth our future.

I ran 2 tasks on a 250 card....
http://setiathome.berkeley.edu/results.php?hostid=6031403&offset=0&show_names=0&state=0&appid=5

with default settings

//TQ

---

TRuEQ & TuVaLu

How do I run 1 AP + 2 MBs and 3 MBs when there's no AP available?

Tried .33 on all CUDA counts, thinking that the switch -instances_per_device 1 would insure only 1 AP. But it started 2 APs, so what function do that switch have?

0.33 on all CUDA counts, and -instances_per_device 3.

---

Boinc....Boinc....Boinc....Boinc....

How do I run 1 AP + 2 MBs and 3 MBs when there's no AP available?

Tried .33 on all CUDA counts, thinking that the switch -instances_per_device 1 would insure only 1 AP. But it started 2 APs, so what function do that switch have?

Started 3 but running only 1. 2 other will await free "slot".
But this will negatively affect on elapsed time of awaiting tasks (boinc thinks they running). And can lead to -177 errors (timeout).

When MultiBeam NV will available, this option will allow fine tune AP/MB proportion for every GPU. BOINC has no such ability at all. It's part of long time ignored ProjectPairing proposal..

Just finished the first unit.

2600 Seconds on a GTX480 vs 60,000 on a Xeon 5130.

http://setiathome.berkeley.edu/results.php?hostid=5955102&offset=0&show_names=0&state=0&appid=5

Hope it validates.

-Darren


[/url]

---

today i had a new one AP -> http://setiathome.berkeley.edu/result.php?resultid=2003222228
only change to the standard settings was unroll 7 which gave to less GPU-load (around 55 %). One nV driver crash happens at the restart-time of the AP. But i had no obvious problems before (no lags, gpu-z data was all ok). In event log, only msg around this time was "Event ID 4101 — Display Driver Timeout Detection and Recovery" (nvlddmkm).
And after that, the calculation continues without further problems (no downclocking for example).

greetings

Chris

---

- Performance is not a simple linear function of the number of CPUs you throw at the problem. -

One day when the sun is shining, and I have tail wind, I will try this app on my tiny ION.

Well, you're Partly in Luck! There's Plenty of Sunshine to be had on this side of the Pond. Starting anywhere on the East Coast to ummm...just about Osaka, or thereabouts! :)

Lt

we'll send you some!

Yes please - its been quite murky over here for the last couple of weeks...

But we don't want it so hot that it fries our low temperature acclimatised anatomical components...

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

Hmmmm, wonder why this one decided to go inconclusive? http://setiathome.berkeley.edu/workunit.php?wuid=781560387

---

PROUD MEMBER OF Team Starfire World BOINC

Just finished the first unit.

2600 Seconds on a GTX480 vs 60,000 on a Xeon 5130.

http://setiathome.berkeley.edu/results.php?hostid=5955102&offset=0&show_names=0&state=0&appid=5

Hope it validates.

-Darren


[/url]

Validated. Just used the rescheduler to move AP units from CPU to GPU. Cross your fingers.

---