[QUOTE=VictordeHolland;355009]There are noticeable speed differences between the bitlevels/kernels as most people may have noticed.

On my HD7950 with settings:
SieveOnGPU=1
GPUSievePrimes=82486
GPUSieveSize=64
GPUSieveProcessSize=16


2^65 - 2^68 gives ~250 GHzdays/day uses cl_barrett32_76 (more bitlevels at once)(horrible on 7xxx cards)(LMH range)
2^68 - 2^69 gives ~405 GHzdays/day uses cl_barrett15_69 (DC TF range)
2^69 - 2^70 gives ~365 GHzdays/day uses cl_barrett15_71 (DC TF range)
2^70 - 2^71 gives ~330 GHzdays/day uses cl_barrett15_73 (DC TF range)
2^71 - 2^72 gives ~330 GHzdays/day uses cl_barrett15_73 (LL TF range)
2^72 - 2^73 gives ~330 GHzdays/day uses cl_barrett15_73 (LL TF range)
2^73 - 2^74 gives ~295 GHzdays/day uses cl_barrett15_82 (LL TF range)[/QUOTE]

Interesting. I'll have to try tuning for both DC and LL.

[SOLVED: the "Power Control Settings" had to be bumped up (currently at +10%) in AMD Overdrive, which is under the Performance tab in AMD's Catalyst Control Center software]

I'm seeing something odd.

Running a Trial Factor from 2^71 to 2^72

My GHzdays/day start out ~350.
About 30 seconds in it drops to ~320.
After another 30 seconds it drops to ~230.

My setup:[LIST][*]Radeon 7950 w/ boost (Sapphire 3L) @ stock clocks[*]AMD Catalyst 13.1[*]Windows 7 64-bit[*]mfakto-0.13: running mfakto_x64 with all default .ini settings except VectorSize=2 for GCN.[/LIST]
Sounds like I should be seeing around 300 GHzdays/day. I've tried changing various settings in the mfakto.ini, downclocking GPU to 800 MHz, clean installs of Catalyst 12.10, 13.1, 13.4, 13.9 and 13.11beta (both of the last two result in display driver resets). Nothing seems to keep me steady at 300 GHzd/d. None of the GPU-z sensors seem to change during the reductions either.

Any ideas?

My guess, thermals. Have you checked clocks, temperatures, etc wile running? It may be throttling.

[QUOTE=kracker;355957]My guess, thermals. Have you checked clocks, temperatures, etc wile running? It may be throttling.[/QUOTE]
Thanks for the reply. I found my issue and updated my message (2 posts up) with the answer. Now sitting at ~350 GHzd/d and steady.

Just for fun I tried compiling mfakto under Windows/Mingw/GCC.

[QUOTE=kracker;358327]Just for fun I tried compiling mfakto under Windows/Mingw/GCC.[/QUOTE]

That's strange ... I thought that the open()/close() file modes are defined by one or another gcc standard header. Either additional headers need to be included, or you'd need to switch to the windows branch of the code (try defining _MSC_VER to something that could mimic a recent MS compiler).

BTW, I managed to reproduce the high-CPU-load issue of the newer AMD drivers using a [URL="http://devgurus.amd.com/message/1301028"]rather simple test program[/URL] (still being moderated) - lets see if AMD has any help to offer. When copying the 13.1 amdocl64.dll into the test program's directory, it runs at 0.04% CPU.

[QUOTE=Bdot;358414]That's strange ... I thought that the open()/close() file modes are defined by one or another gcc standard header. Either additional headers need to be included, or you'd need to switch to the windows branch of the code (try defining _MSC_VER to something that could mimic a recent MS compiler).

BTW, I managed to reproduce the high-CPU-load issue of the newer AMD drivers using a [URL="http://devgurus.amd.com/message/1301028"]rather simple test program[/URL] (still being moderated) - lets see if AMD has any help to offer. When copying the 13.1 amdocl64.dll into the test program's directory, it runs at 0.04% CPU.[/QUOTE]

Good, I'll be very happy when they fix those two :smile:

After a long time of nothing... I finally got around to trying it again.

After a ton of warnings:

mfakto 0.14pre2 - for testing only
 

I can't access mersenneforum.org/mfakto at the moment, so I attached it here. Please give this one a try. Most important thing is that it should not abort nor make Catalyst die anymore. Second, when GPU-sieving, it reads a new config setting, FlushInterval. It determines after how many scheduled kernels it will wait for completion. Low values (1 .. 3) reduced the CPU load to almost zero, but lower values generally have a slightly lower throughput. 0 is unlimited which was the old behavior.

It also has a fix for a small memory leak (~0.5kB per assignment).

It is not yet completely tested - do not yet submit no-factor results to primenet.

Ah, and it tries to read a config setting, UseBinfile, which is not yet fully implemented - please ignore.

Will do :smile: I have 13.4 here, will test other computer with 13.10 later this weekend

EDIT:

AMD Catalyst 13.4

0.14p2=0.3-0.5% cpu, basic selftest pass.
0.13=crash

Catalyst 13.11 beta 9.4. Test case 2966/32927 for st2 and no failures so far. Looking good. :tu: