SkylakeX teasers (aka prime95 29.5)

[Prime95]

Originally Posted by ET_

Is version 29.5.10 able to recover from such situation?
What should I do with the *.bad savefiles?

There is nothing new for you in build 10.

You can try to exit prime95, rename one of the .bad biles to .bu, restart prime95. There is the tiniest chance that the save files were good and the program or OS was in a funky state. I've seen it happen in the rarest of occasions.

Otherwise, delete the save files. Run a torture test. Restart from scratch. Increasing the frequency of Jacobi checks might be prudent until you are sure the system is stable.

[mackerel]

Just tried using 29.5b10 for some benching.

First an observation, it doesn't seem to remember the number of workers in the benching settings window, where other settings are remembered. On 8086k and 7800X systems, it would always go back to 1, 6 when I return to the window, having previously entered only 6 in there. On my R5 2600 system, it reverts to 1, 2, 6, even though it is 6 core 12 thread like the Intel CPUs. I used a fresh download of P95 so I don't think it should be affected by any past testing with older versions.

I also tried using it to see how different instructions affect performance by editing local.txt according to undoc.txt. In short, I see 4 performance levels, let's call them AVX512F, FMA3, AVX, SSE/SSE2/SSE4. Is AVX2, distinct from FMA3, used at all? Disabling AVX2 in isolation doesn't seem to make any difference.

I think in older versions the output would show which instruction or FFT type (not size) was in use. I don't see it any more, has it been removed?

[Prime95]

Originally Posted by mackerel

First an observation, it doesn't seem to remember the number of workers in the benching settings window, where other settings are remembered. On 8086k and 7800X systems, it would always go back to 1, 6 when I return to the window, having previously entered only 6 in there. On my R5 2600 system, it reverts to 1, 2, 6, even though it is 6 core 12 thread like the Intel CPUs.

Yeah, it defaults to the #workers a newbie should try to see what #workers gives the best throughput. I'm not sure I like that feature either.

12 is not one of the defaults because using hyperthreading is often a poor choice.

Is AVX2, distinct from FMA3, used at all? Disabling AVX2 in isolation doesn't seem to make any difference.

The integer instructions introduced by AVX2 are used in trial factoring. You are correct the AVX2 flag makes no difference in the FFTs.

I think in older versions the output would show which instruction or FFT type (not size) was in use. I don't see it any more, has it been removed?

Look in results.bench.txt. I don't think the FFT type has ever been displayed on screen.

[PhilF]

Originally Posted by Prime95

Look in results.bench.txt. I don't think the FFT type has ever been displayed on screen.

On my i7-4790, running Prime95 V29.4 build 8, the screen output shows it is using "FMA3 FFT length 4800K". It's running a PRP test, not a LL test.

[Prime95]

Originally Posted by PhilF

On my i7-4790, running Prime95 V29.4 build 8, the screen output shows it is using "FMA3 FFT length 4800K". It's running a PRP test, not a LL test.

Clarification: I don't think it has ever shown the "FMA3" on screen during a benchmark.

[simon389]

79207379 PRP DC on build 10 on the still “broken” system completed successfully!

Next up is 79210559.

[simon389]

I'm still having major hardware troubles with the 29.5b10 release. I was able to successfully get it stable with AVX512 tests on AIDA64 (all I needed to do was keep the RAM below 3600 Mhz), but when I moved over to testing on Prime95, I ran into major problems.



29.4 still works great on all 3 9800X machines, but 29.5b10 cannot get a stable LL doublecheck, no matter what settings I use.

I've set the CPU from 3.5 Ghz to 4.1 Ghz - all have hardware errors in LL DC. I've set the RAM from its rated 3600Mhz speed to motherboard default 2000Mhz (as well as slowed the cas latency and other RAM timings, even at 2000Mhz). As a last resort I tried a severely underclocked 3.6 Ghz with increased CPU voltage (1.15v) and 2000Mhz RAM with slower CL19 timings. All on a 850W PSU, Mhz/voltages monitored with CPU-Z and temps monitored. And this failed too.



So I'm at a loss for what to try next, and feel this 29.5 release is still going to be problematic for some AVX512 CPUs. Unless I'm missing some setting to try.

Feel free to suggest I move to another thread, because perhaps my systems are just unique.

[simon389]

So this is peculiar. I am three hours into a blend torture test and everything looks great (perhaps temps get a bit high at times - 85C). But if I try to run a LL DC it very quickly errors out with "Invalid FFT, restarting from save file". So the AVX512 torture test seems like it's working but the actual LL DC seems like its erroring. Gives me hope that maybe the problem is the beta code and not my hardware.

[GP2]

Originally Posted by simon389

Gives me hope that maybe the problem is the beta code and not my hardware.

But others are using the same code and getting matching LL double-checks and PRP double-checks.

In my case I'm using the Linux version so it's not quite the same, but surely many have used the Windows versions of the various 29.5 builds.

Maybe you could try running the Linux 64 version under WSL on the very same machines, and see what happens.

Or double-check prime.txt and local.txt to see if there's some unusual setting in there that no one else uses.

[VBCurtis]

Originally Posted by simon389

So this is peculiar. I am three hours into a blend torture test and everything looks great (perhaps temps get a bit high at times - 85C). But if I try to run a LL DC it very quickly errors out with "Invalid FFT, restarting from save file". So the AVX512 torture test seems like it's working but the actual LL DC seems like its erroring. Gives me hope that maybe the problem is the beta code and not my hardware.

Have you tried moving the save file out of the folder and starting a test from scratch now that the hardware is more stable? Perhaps your save file is part of the problem? (Says the guy who doesn't know quite how the Gerbicz error check works, but has paid enough attention to have read that your machine somehow had an error despite the check)

[simon389]

Originally Posted by VBCurtis

Have you tried moving the save file out of the folder and starting a test from scratch now that the hardware is more stable? Perhaps your save file is part of the problem? (Says the guy who doesn't know quite how the Gerbicz error check works, but has paid enough attention to have read that your machine somehow had an error despite the check)

Yes, I've tried starting from scratch many times. My Prime95 stress test (blend) is now going on 11 hours with no errors, but within 5 minutes of starting a LL DC it still errors Invalid FFT. So strange...

Originally Posted by GP2

But others are using the same code and getting matching LL double-checks and PRP double-checks.

In my case I'm using the Linux version so it's not quite the same, but surely many have used the Windows versions of the various 29.5 builds.

Maybe you could try running the Linux 64 version under WSL on the very same machines, and see what happens.

Or double-check prime.txt and local.txt to see if there's some unusual setting in there that no one else uses.

I've always used brand new installations of build 10, so I'm unsure how prime.txt and local.txt could have bad settings. I will install Linux, run Prime95 via Linux, and see if that helps.