Prime95 version 27.6

[Dubslow]

Originally Posted by BrianOC

Run 8 theards and on SUM(INPUTS) erorr checking and Round off error checking in both.

When you say both, you mean in both 27.6 and 27.4, or do you mean in both 27.6.2 and 27.6.3? (And that's fine about your English, we can (mostly) understand enough, or at least I can )

[BrianOC]

Originally Posted by Dubslow

When you say both, you mean in both 27.6 and 27.4, or do you mean in both 27.6.2 and 27.6.3? (And that's fine about your English, we can (mostly) understand enough, or at least I can )

Test 3. (27.6.3) Crash again same time 2hour30 minutes about.

Image: http://i48.tinypic.com/2lt2bmo.png

passed 10 run 11 crashed.
[Mon Apr 30 14:18:05 2012] start, [Mon Apr 30 16:36:04 2012] crash

I decided for the time being using 27.4 stable

[Dubslow]

Originally Posted by BrianOC

Test 3. (27.6.3) Crash again same time 2hour30 minutes about.

Image: http://i48.tinypic.com/2lt2bmo.png

passed 10 run 11 crashed.
[Mon Apr 30 14:18:05 2012] start, [Mon Apr 30 16:36:04 2012] crash

I decided for the time being using 27.4 stable

Hmm... now crashes as well as errors. I really don't know. George? I'll probably start another blend torture test and let it run for a few hours, but I haven't had any crashes doing production work (though I'm in Linux of course, not Windows).

[BrianOC]

Originally Posted by Dubslow

Hmm... now crashes as well as errors. I really don't know. George? I'll probably start another blend torture test and let it run for a few hours, but I haven't had any crashes doing production work (though I'm in Linux of course, not Windows).

Blend no crash in 6h44 min. Small FTT error always the same time/place.

[Dubslow]

Originally Posted by BrianOC

Blend no crash in 6h44 min. Small FTT error always the same time/place.

Yeah, I just ran blend for five hours, but have now started Small FFT testing. That time is roughly 2.5 hours into the test, right?

[BrianOC]

Originally Posted by Dubslow

Yeah, I just ran blend for five hours, but have now started Small FFT testing. That time is roughly 2.5 hours into the test, right?

yes

[Dubslow]

I'm pretty sure a seg fault is code error

[Worker #1 Apr 30 17:44:52] Test 27, 160000 Lucas-Lehmer iterations of M630783 using AVX FFT length 50K, Pass1=640, Pass2=80.
[Worker #3 Apr 30 17:45:06] Test 26, 160000 Lucas-Lehmer iterations of M638977 using AVX FFT length 50K, Pass1=640, Pass2=80.
[Worker #2 Apr 30 17:45:14] Test 28, 160000 Lucas-Lehmer iterations of M622593 using AVX FFT length 50K, Pass1=640, Pass2=80.
[Worker #1 Apr 30 17:45:30] Test 28, 160000 Lucas-Lehmer iterations of M622593 using AVX FFT length 50K, Pass1=640, Pass2=80.
[Worker #3 Apr 30 17:45:43] Test 27, 160000 Lucas-Lehmer iterations of M630783 using AVX FFT length 50K, Pass1=640, Pass2=80.
[Worker #2 Apr 30 17:45:50] Test 29, 160000 Lucas-Lehmer iterations of M614399 using AVX FFT length 50K, Pass1=640, Pass2=80.
[Worker #1 Apr 30 17:46:07] Test 29, 160000 Lucas-Lehmer iterations of M614399 using AVX FFT length 50K, Pass1=640, Pass2=80.
[Worker #3 Apr 30 17:46:20] Test 28, 160000 Lucas-Lehmer iterations of M622593 using AVX FFT length 50K, Pass1=640, Pass2=80.
Segmentation fault

Again, all single threaded, MPrime-64, 2600K; it's build 3, but when I run -v I get 27.6.8.

Edit: @Jwb52z: Yeah, I'm starting to doubt it, but I haven't yet encountered any problems on my desktop doing production work, and it has continued to turn in good DCs, most recently around 10-12 hours ago; my laptop ran into the first round off bug, but admittedly has been borked the last few days.

[Jwb52z]

These last few posts almost make me glad I haven't started using 27.6 yet.

[Prime95]

Originally Posted by Dubslow

it's build 3, but when I run -v I get 27.6.8.

The ".8" refers to the OS port # -- a pretty meaningless number for end-users. I'll update mprime to output the more human-readable string that prime95 uses and mprime uses when talking to the server.

[Prime95]

Originally Posted by BrianOC and others

Blend no crash in 6h44 min. Small FTT error always the same time/place.

At first blush it looks like a problem in the 16K FFT (actually the carry propagation code) when the exponent gets too small (only 10.5 bits per FFT word). This would never happen in the real world unless one explicitly requests using a longer FFT length than the default.

Nevertheless, it is a bug and I'll track it down. Thanks for finding it.

[BrianOC]

Originally Posted by Prime95

At first blush it looks like a problem in the 16K FFT (actually the carry propagation code) when the exponent gets too small (only 10.5 bits per FFT word). This would never happen in the real world unless one explicitly requests using a longer FFT length than the default.

Nevertheless, it is a bug and I'll track it down. Thanks for finding it.