New AstroPulse for GPU ( ATi & NV) released (r1316)

I tell you here when downloading my 2 units GPU is finally completed ...
____________


Alliance Francophone

About 1 hour 20 for the operating time on my gtx 470
____________


Alliance Francophone

I've noticed that some WUs are getting low credits, is there a problem somewhere? It's not happening only with the latest app, it also happened with previous versions, not sure when I first noticed it.

http://setiathome.berkeley.edu/results.php?hostid=6442255&offset=0&show_names=0&state=3&appid=12

About 1 hour 20 for the operating time on my gtx 470

I had an abortion because the download is very long ...

In my preferences I put yes to everything in CPU / GPU

I had an abortion because the download is very long ...

In my preferences I put yes to everything in CPU / GPU

Well, DownLoads are very slow at SETI, comaired to projects with
fewer users and hosts, GPUgrid hits 20MBit, Einstein is also faster.
Milkyway, WCG, etc..
Problem is SETI is constant overloaded/saturated and in too much cases
the cause of a DownLoad Errors as in your case.

How long for a MB (~360KBytes) and AstroPulse (8MBytes) workunit?

Just some more feedback for Raistmer and any other AP OpenCL developers:

Getting a lot of invalid results on my old HD 4670, mostly due to 30 single pulses, 30 repeating pulses, or both. This seems to happen regardless of what combination of parameters I try. The confusing thing is that, sporadically, I still get the occasional valid result.

I think I'll have to revert to r555 even though it has much lower performance (especially with regards to CPU usage), it's still bound to be much better than CPU-only processing on the ancient machine it's on. There were still invalid results, but they were limited to 30 repeating pulses and there were far more valid results than invalid ones. I put all this down to the 'beta-level' OpenCL support on the HD 4000 series - I ceased all GPU processing on my HD 4850 a long time ago because of similar issues presumably arising from this.

Other GPUs (HD 5000 series and later, as well as GeForce cards) continue to work very well on r1316. I'm aware there is a r1363 release, but I don't understand the parameters enough to try it. Doesn't seem as though there's as large a performance increase either.
____________
Soli Deo Gloria

Have you tried lower unroll ?

I´d suggest -unroll 4 -ffa_block 2048 -ffa_block_fetch 1024

Mike

____________

Yes. As I mentioned, the both valid and invalid results occur for all manner of parameters. I've tried unroll of 4, 8, 10 with r1316. With r555, unroll of 4 always resulted in invalid results, whereas 10 was the only one that consistently returned valid results.
____________
Soli Deo Gloria

I would be delighted to just download some AstroPulse files that would run! Every one shows completion in the Transfers tab, but the Tasks list always says 'Download failed'.

Running all other WU's in Seti and Einstein with no problems.

Windows 7 Pro 64-bit; i7-2600 quad core with 8 Gb RAM.
____________

Just some more feedback for Raistmer and any other AP OpenCL developers:

Getting a lot of invalid results on my old HD 4670, mostly due to 30 single pulses, 30 repeating pulses, or both. This seems to happen regardless of what combination of parameters I try. The confusing thing is that, sporadically, I still get the occasional valid result.

I think I'll have to revert to r555 even though it has much lower performance (especially with regards to CPU usage), it's still bound to be much better than CPU-only processing on the ancient machine it's on. There were still invalid results, but they were limited to 30 repeating pulses and there were far more valid results than invalid ones. I put all this down to the 'beta-level' OpenCL support on the HD 4000 series - I ceased all GPU processing on my HD 4850 a long time ago because of similar issues presumably arising from this.

Other GPUs (HD 5000 series and later, as well as GeForce cards) continue to work very well on r1316. I'm aware there is a r1363 release, but I don't understand the parameters enough to try it. Doesn't seem as though there's as large a performance increase either.

Thanks BilBg. Trend did try to block it, and I placed it on the 'approved' list. I replaced the one in the folder which was only 9kB. Time will tell, once I get another AP download.
____________

Special LHD4 build for them still in debugging. From the other side I would not expect same failures on HD4850. Can you replace HD46xx with HD48xx on that host ?

Edit: If you want a tester for your special HD 4000 build, I'm willing to help.

I'm in no hurry, I only offered because you mentioned developing a special build.

Best wishes for MB development.
____________
Soli Deo Gloria

AstroPulse WU's are running fine again. :)
____________

Only as Info, Application runs without errors on a low low lowest end HD4350 ^^ ~24h for 1 WU
____________