Skylake FMA3 round off error

[Madpoo]

Originally Posted by LaurV

Yes, Madpoo can easily do that, after he clear out all my bad results I had between January and March 2012 when I was testing the new cudalucas (switching from powers-of-two-FFT to non-powers-of-two-FFT). Didn't I say that many times?

The rule was that the bad results stay bad. For whatever reason. Otherwise we completely mess the statistics... Are we doing statistics or not?

That's my take on it.

I even have a few bad ones that were the result of a Prime95 bug (shift count being smaller than the exponent itself). I won't lose sleep over it. They're still in there.

[tha]

I saw there is an option on the detailed stats page of each account to 'reset' the individual processor's reputation once a hardware repair has been made. So I executed that option.

[ATH]

Your result was bad unfortunately, I hope you fixed the problem:

http://mersenne.org/M43175681

[tha]

I also tested exponent 43175683. First time I had the cover taken off. Unfortunately someone tried to be helpful and tried to fit the cover back on, even a quarter turned, and using some extra force. Unsuccessfully, but exerting some extra pressure on the motherboard. The result was a LL residue that didn't match the result in the database.
Next, I modified the GPU power cable so that it became much more flexible by taking off the tie wraps and shrink tubing. Now the cabinet could be properly closed and I restarted the exponent from zero.
The second run produced a result that did match original result in the database,

However, when I look up my results I get this:

Skylake-1	43175683	C - Bad	2016-02-06 11:08	1.0	8A6C129285C1F8__	65.9029
Riet-Ubuntu	13695961	NF-PM1	2016-02-06 08:31	0.0	B1=300000, B2=6000000	0.4240
Riet-Ubuntu	13695811	NF-PM1	2016-02-06 01:34	0.0	B1=300000, B2=6000000	0.4240
Riet-Ubuntu	13694831	F-PM1	2016-02-05 18:32	0.0	1169316217401148086056503	0.4240
Manual testing	11971457	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11737441	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11971637	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11737367	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11971537	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11737301	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11970709	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11736943	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11970169	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11736727	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11970017	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11736559	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11969987	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11736577	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11969723	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11736113	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11969701	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11735813	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11969329	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11735783	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11969323	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11734907	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Manual testing	11735177	NF-PM1	2016-02-05 12:28	0.0	B1=300000, B2=6000000	0.3408
Riet-Ubuntu	13694377	NF-PM1	2016-02-05 11:30	0.0	B1=300000, B2=6000000	0.4240
Riet-Ubuntu	13693117	NF-PM1	2016-02-05 04:28	0.0	B1=300000, B2=6000000	0.4240
Skylake-1	43175683	C - Bad	2016-02-05 03:16	30.2	DDA77236747D6E__	65.9029
Riet-Ubuntu	13712833	NF-PM1	2016-02-04 21:26	0.0	B1=300000, B2=6000000	0.4240
Manual testing	15996941	NF-PM1	2016-02-04 16:39

That is a bit weird, the latest result should be listed as correct.

[ATH]

Your 2nd result did match: http://mersenne.org/M43175683

[tha]

Originally Posted by ATH

Your 2nd result did match: http://mersenne.org/M43175683

It does, but as in my previous post, it is listed as 'BAD' on my stats page. In the GIMPS database the exponent is correctly cleared as valid double check. The error may be because the software does not expect two separate instances of a LL test by one user with different results, or because of anything else.

Furthermore I found out that the option 'FullBench=1' is still working if added to 'prime.txt'. That option is not listed in 'undoc.txt' (anymore). If you add that line and choose option '16 Benchmark' in mprime or just 'Benchmark' from the prime95 GUI, benchmark will list each and every available FFT size and timing.
It becomes obvious that some sizes are slower than the next bigger FFT size.

[ATH]

Originally Posted by tha

It becomes obvious that some sizes are slower than the next bigger FFT size.

That happens a lot on CUDALucas but only for 1 FFT for me in Prime95.

I do not think you can trust that benchmark too much for that as it does only 25-50 iterations for each FFT, and it gave me some slightly weird results.

Instead I started real LL tests with:
UsePrimenet=0
PercentPrecision=6
TimingOutput=4
in prime.txt. I tested 100k-150k iterations of each FFT with screen output every 25k so I got at least 3-4 screen outputs from each FFT. I think screen output every 10k would have been fine though and doing 40-50k iterations total.

I tested FFT 1920K (36M exponents) up to 4800K FFT (90M+) with and without "ErrorCheck=1" and "SumInputsErrorCheck=1" in prime.txt, whch I normally use. Last time I did a small test on errorcheck I found it "cost" 0.6%-1.2% extra time, but it turned out to be more now with a more thorough test: 0.8%-2.8% extra.

The only FFT size I should avoid is 2240K which is slower than 2304K.

				No Errorcheck	Errorcheck
				ms/iter		ms/iter
FMA3 1920K	M36000007	1.161-1.166	1.191-1.192 (+2.4%)	
FMA3 2048K (2M) M38000009	1.238-1.238	1.265-1.269 (+2.3%)	
FMA3 2240K	M40000003	1.405-1.405	1.436-1.440 (+2.3%)	
FMA3 2304K	M43500011	1.363-1.363	1.392-1.401 (+2.5%)	
FMA3 2400K	M44500021	1.479-1.486	1.517-1.525 (+2.6%)	
FMA3 2560K	M46500017	1.574-1.579	1.600-1.608 (+1.7%)	
FMA3 2688K	M49500023	1.668-1.669	1.695-1.699 (+1.7%)	
FMA3 2880K	M52000019	1.821-1.825	1.843-1.848 (+1.2%)	
FMA3 3072K (3M)	M56000029	1.915-1.917	1.963-1.976 (+2.8%)	
FMA3 3200K	M59000041	2.035-2.035	2.075-2.082 (+2.1%)
FMA3 3360K	M61500059	2.189-2.189	2.210-2.222 (+1.2%)	
FMA3 3456K	M64500049	2.281-2.281	2.309-2.317 (+1.4%)	
FMA3 3584K	M67000033	2.336-2.337	2.367-2.376 (+1.5%)	
FMA3 3840K	M69000047	2.505-2.507	2.531-2.545 (+1.3%)	
FMA3 4096K (4M)	M73500001	2.775-2.778	2.801-2.820 (+1.2%)	
FMA3 4608K	M78500027	3.092-3.097	3.115-3.126 (+0.8%)	
FMA3 4800K	M88000061	3.267-3.272	3.311-3.324 (+1.5%)