AP on Nvidia GPU - Am I Doing it Right ?

[Terror Australis]

I've just gotten a test machine machine crunching as an AP only rig. It's processing but with SAH down I can't see if it's producing valid results yet, so if someone experienced can look at these figures and give me a "yes, no, maybe" I'd appreciate it.

The machine is as follows. QX9650 + GTX470. Win XP 32bit, 266.58 Driver. AP6_win_x86_SSE_CPU_r1797 CPU app, AP6_win_x86_SSE2_OpenCL_NV_r1761 GPU app.

My ap_cmdline file is straight out of the readme at "-unroll 12 -ffa_block 8192 -ffa_block_fetch 4096 -hp".

 Average GPU crunching time after 4 units is around 35 minutes, CPU usage is between 1 and 3%. According GPUZ, GPU load is 98% and the Memory Controller Load is a square wave varying between 12 and 80% on a quite regular basis.

The BOINC ETA counter counts up for around 10 seconds and then suddenly drops by around 10 minutes and the cycle starts again.

CPU crunching time is around seven and a half hours.

Do these figures look right for this setup ?

Currently have one core of the CPU set to feed the GPU but the usage is so low that the System Idle Process is running nearly 25% so I'll probably re-enable the 4th core.

Thanks to Claggy and Mike for their help on the SAH boards.

T.A.

[Raistmer]

with this old driver you can try to load all CPU cores.
Other params look right.

[Claggy]

The machine is as follows. QX9650 + GTX470. Win XP 32bit, 266.58 Driver. AP6_win_x86_SSE_CPU_r1797 CPU app, AP6_win_x86_SSE2_OpenCL_NV_r1761 GPU app.

My ap_cmdline file is straight out of the readme at "-unroll 12 -ffa_block 8192 -ffa_block_fetch 4096 -hp".

 Average GPU crunching time after 4 units is around 35 minutes, CPU usage is between 1 and 3%. According GPUZ, GPU load is 98% and the Memory Controller Load is a square wave varying between 12 and 80% on a quite regular basis.

35 minutes is what i get with my factory overclocked 460GTX with -unroll 10 -ffa_block 6144 -ffa_block_fetch 1536 and recent drivers,
you'd probably get a bit of a speedup too, but with loss of CPU productivity from saving a CPU core, hard to tell what's best.

Claggy

[Mike]

In addidtion, if you get units with higher blanking you will need the CPU core probably because blanking is calculated on CPU only.

Or you need to investigate your run times carefully.
On following WU start  up the app can go in stall without a core freed.
Depends how many instances you are running.

[Raistmer]

He runs older driver that do not have these issues with all cores busy.

[Terror Australis]

An update.
I changed the driver to v310.70.

It appears to have made no, or very little difference to the GPU crunching times compared to the earlier version.

What parameters in the ap_cmdline file are the ones most likely to give an improvement if tweaked ? And which would be the best way to adjust them, up or down ?

As stated above, this is a dedicated AP only cruncher so it can be tweaked for maximum performance without worrying about the effects on other programs.

T.A.

Edit: Just for information
After 5 units each, it does seem that, using the same settings as in my op, the v310.70 drivers are 15 to 30 secs slower than the v266.58 edition.

[Mike]

We need more details to give good advise.
How many instances are you running ?

8192 4096 is a good value which i`m ussing for 2 years now running 2 instances.
Maybe 12288 6144 is possible but i wouldn`try it whilst seti is offline to be honest.

[Terror Australis]

We need more details to give good advise.
How many instances are you running ?

8192 4096 is a good value which i`m ussing for 2 years now running 2 instances.
Maybe 12288 6144 is possible but i wouldn`try it whilst seti is offline to be honest.

Only one ATM. I'm still experimenting.

T.A.