Prime95 v30.4/30.5/30.6

[Chuck]

Originally Posted by ATH

I guess this could not be an mprime issue? Just in case could someone with a Prime95 (Windows) AVX-512 computer test one of these? No need to finish it unless you want to, just run like 50% of stage1 to check if you get a roundoff error:
Pplus1=N/A,1,2,4037479,-1,2500000,250000000,2,70
Pplus1=N/A,1,2,4040059,-1,2500000,250000000,2,70

I'll try these two.

[Chuck]

I didn't get any roundoff errors.

[Citrix]

If doing P-1 work on (k^p-1)/(k-1) or (k^p+1)/(k+1) with p prime
Does Prime95 automatically includes p in the B1 stage?


Thanks

[ATH]

Originally Posted by Chuck

I didn't get any roundoff errors.

@Prime95: Is this strange that there are no roundoff errors in Prime95 on AVX-512 200K on these 2 exponents that I got roundoff errors on in mprime? or is it very dependent on the specific processor?

[Prime95]

Originally Posted by Citrix

If doing P-1 work on (k^p-1)/(k-1) or (k^p+1)/(k+1) with p prime
Does Prime95 automatically includes p in the B1 stage?

I believe so. Can you run a small test case?

[Prime95]

Originally Posted by ATH

@Prime95: Is this strange that there are no roundoff errors in Prime95 on AVX-512 200K on these 2 exponents that I got roundoff errors on in mprime? or is it very dependent on the specific processor?

One started at 2/7, the other started at 6/5.

[ATH]

Originally Posted by Prime95

One started at 2/7, the other started at 6/5.

Yeah I switched it :( In case whoever tested it wanted to finish it, then it would not be wasted effort.

Originally Posted by Prime95

Other AVX-512 FFT maximum exponents that remain to be tested (turn on roundoff checking):

22469
33423

I'm working on these 2 and going upwards after them. I'm doing P+1, do I need to do P-1 as well on each exponent?
Should I continue above the maximum in case of no failures below?

Edit: I reached a failure in the 33423 case. I started at M32611 and reached a roundoff error at M33391, should I do more testing for this FFT?

[Work thread Jun 28 16:01:03] P+1 on M33391, start=2/7, B1=100000000, B2=10000000000
[Work thread Jun 28 16:01:03] Using AVX-512 FFT length 1536
[Work thread Jun 28 16:09:30] Possible roundoff error (0.4520854), backtracking to last save file.
[Work thread Jun 28 16:09:30] Using AVX-512 FFT length 1536
[Work thread Jun 28 16:17:55] Possible roundoff error (0.4520854), backtracking to last save file.
[Work thread Jun 28 16:17:55] Using AVX-512 FFT length 1536
[Work thread Jun 28 16:26:17] Possible roundoff error (0.4520854), backtracking to last save file.

Max roundoff errors before the failure:
M33301 Round off: 0.4065263058
M33347 Round off: 0.4113334932
M33353 Round off: 0.4064540519
M33359 Round off: 0.401012083
M33377 Round off: 0.4146345477

[R. Gerbicz]

Originally Posted by ATH

Yeah I switched it :( In case whoever tested it wanted to finish it, then it would not be wasted effort.
...
I'm doing P+1, do I need to do P-1 as well on each exponent?
Should I continue above the maximum in case of no failures below?

Edit: I reached a failure in the 33423 case. I started at M32611 and reached a roundoff error at M33391, should I do more testing for this FFT?

[Work thread Jun 28 16:01:03] P+1 on M33391, start=2/7, B1=100000000, B2=10000000000
[Work thread Jun 28 16:01:03] Using AVX-512 FFT length 1536
[Work thread Jun 28 16:09:30] Possible roundoff error (0.4520854), backtracking to last save file.
[Work thread Jun 28 16:09:30] Using AVX-512 FFT length 1536
[Work thread Jun 28 16:17:55] Possible roundoff error (0.4520854), backtracking to last save file.
[Work thread Jun 28 16:17:55] Using AVX-512 FFT length 1536
[Work thread Jun 28 16:26:17] Possible roundoff error (0.4520854), backtracking to last save file.

Max roundoff errors before the failure:
M33301 Round off: 0.4065263058
M33347 Round off: 0.4113334932
M33353 Round off: 0.4064540519
M33359 Round off: 0.401012083
M33377 Round off: 0.4146345477

For example on M33391 we know a pretty large d (=p1*p2*p3*p4) divisor which enables a good error checking at least on the first stage of P+-1 : do the computation also in Z[d] and then regularly do the check for the intermediate number(s) that r(i) mod d is still good. What could give you a greedy FFT size.

Or for example at P-1 test the check is only: we computed that b^e==res mod N then check: res==b^(e mod eulerphi(d)) mod d [Euler-Fermat], for gcd(b,d)=1 and d|N, so you need to update only the e mod eulerphi(d) at each step and not the slightly more costly b^e mod d. [since d is usually very small compared to N this makes not a large difference in timing].

[Prime95]

Originally Posted by ATH

I'm doing P+1, do I need to do P-1 as well on each exponent?
Should I continue above the maximum in case of no failures below?

P+1 is sufficient. Don't go above the maximum if no failures are found.

I started at M32611 and reached a roundoff error at M33391, should I do more testing for this FFT?

To be safe, length 1.5K AVX-512 FFT max exponent changed to 33380.

[Viliam Furik]

I am having a worker-count related issue with Prime95. I have been using (4 workers, 3 cores) configuration for a while. Now I've decided to make it (12 workers, 1 core), because of the better throughput at 50K FFT, which is what I am doing now - P-1 in 1M range on factored exponents. So I went and set the Prime95 accordingly. When I started the workers, it went fine, but when it got to starting worker #6, it crashed. The same happened when I restarted the program.

Workers #1 to #4 have plenty of work in their worktodo entries, worker #5 was assigned CERT, but the rest is hungry. Is it possible that it crashes because of that?

I am using version 30.6b4, processor is Ryzen 9 3900X (12 cores).

After a few testing attempts, it seems to crash precisely when all the first 4 workers finally agreed on RAM consumption (3 of them are in stage 2) and started working. That happens somewhere around the start of worker #8, i.e. about 40 seconds after the start of the program (there are 5-second delays between worker starts).

[Viliam Furik]

Weird... When I fed the workers with assignments, it's no longer crashing.

But I am going to (6 workers, 2 cores), because 12 workers make the CPU really hot, 92 °C.