Strategic Double Clicking

[Madpoo]

Originally Posted by ATH

We have verified 6,575 exponents in this thread over the years (including a few from the Triple Check thread):

verified.html
verified.txt

Hmm... interesting list. Let's see what kind of data I can get from those...

• 13 of those got some extra factoring and a factor was found and only the one test was ever done
• 41 of them had a factor found (so 28 of those already had 2 tests and then a factor was discovered)
• 3091 only have 2 results so they were a match on the 2nd run
• 3353 have a bad result listed so I'm just assuming we must have mismatched the first test and hopefully our 2nd test was correct once a 3rd test was done
• 15 had 5 LL tests done
• 129 had 4 LL tests done (+2 more where a 4th test was for a factor found)
• 3316 had 3 LL tests done (+9 more where a 3rd test was for a factor found)
• 8 had 3 bad results
• 93 had 2 bad results

Don't bother adding any of those #'s together to arrive at any sums... there are some overlaps here and there for different categories. But it's definitely interesting that of the ones that are now verified, we can see that over half of them were worth doing a strategic double-check since they did in fact end up needing at least a 3rd test done.

What I did *not* do was check to see if the verification run matched the first or second test. I think there were probably a few times when our SDC run was bad, but in general I think those were rare. To crunch that part of it I'd have to analyze by the date of each result and I didn't feel like putting in that effort right now. LOL

[ATH]

Yeah we did some good work here

I'm glad I started that "Triple Check" thread back in 2012 (or "Tripple Check" as I mistakenly named it), which later in 2015 inspired you to start triple checking, and then to start this thread.

[LaurV]

Originally Posted by Madpoo

Hmm... interesting list.

Originally Posted by ATH

Yeah we did some good work here

Good job indeed!

[LaurV]

I got assigned 49404263 for DC, and I got a mismatch, but my residue was "funny" in a way... (it seems each hex digit is... doubled, haha), and the first test was done by Chuck Norris curtisc, who doesn't make mistakes, however this test I did with a card which never gave an error up to now, so I got extremely excited and I ran it again in a different card, and guess what, got the same "funny" residue .

I reported both DC and TC, and most probably (like in 99.99999%) this is the right residue, but I know Madpoo will not be happy with the self-DC (in fact, TC) so somebody may do a QC if she likes...

[ATH]

Originally Posted by LaurV

I got assigned 49404263 for DC, and I got a mismatch, but my residue was "funny" in a way... (it seems each hex digit is... doubled, haha), and the first test was done by Chuck Norris curtisc, who doesn't make mistakes, however this test I did with a card which never gave an error up to now, so I got extremely excited and I ran it again in a different card, and guess what, got the same "funny" residue .

I reported both DC and TC, and most probably (like in 99.99999%) this is the right residue, but I know Madpoo will not be happy with the self-DC (in fact, TC) so somebody may do a QC if she likes...

I will start it in 24h after my current exponent finishes, but I cannot reserve it in primenet, as it is already double checked.

[storm5510]

I have 'mooched' the following to do a hardware test:

DoubleCheck=46662349,73,1

I will not submit this for credit!

Reason: When I first installed this 1080, it reported that it would need roughly four days to complete an exponent of this magnitude.

During a recent cleaning, I discovered that the power connector from the PSU was not seating properly, and I was unable to make it do so. I switched to another and it seated with a very clear 'click.' Inspection of the original connector revealed it was incomplete. No latch, and some rounded corners.

Now, instead of four days, it's 30 hours. Something was not making good contact, or any contact at all. I count myself very fortunate that the card was not damaged.

Basically, I want to see if the result I get matches anything already submitted. Nothing further.

[ATH]

It sounds very fast for a 1080 to finish a 46M in 30 hours, but I hope for you that it is correct.

My Titan Black does it in roughly 22 hours, but it should have roughly 1700 GFLOPS of double precision performance while your 1080 only has ~ 250 GFLOPS, because it only has 1/32th of the single precision performance.

[kladner]

30 hours is in the same zone, though a little shorter than an i7-6700K with all cores, one worker. (@4200 MHz cpu, 3200MHz RAM.)

[storm5510]

Originally Posted by ATH

It sounds very fast for a 1080 to finish a 46M in 30 hours, but I hope for you that it is correct.

My Titan Black does it in roughly 22 hours, but it should have roughly 1700 GFLOPS of double precision performance while your 1080 only has ~ 250 GFLOPS, because it only has 1/32th of the single precision performance.

This one is not typical of the specs I've looked at. It's a Gigabyte. I've seen it go above 1100 GHz-d/Days running trial factoring for PrimeNet. Its base clock is 1835 and boosts to 1885. It has gone up to 1915. No overclocking, ever. It does this on its own.

We're a bit off-topic a bit here so I will stop.

[ATH]

Originally Posted by storm5510

This one is not typical of the specs I've looked at. It's a Gigabyte. I've seen it go above 1100 GHz-d/Days running trial factoring for PrimeNet. Its base clock is 1835 and boosts to 1885. It has gone up to 1915. No overclocking, ever. It does this on its own.

We're a bit off-topic a bit here so I will stop.

Yes, I know 1080 is fast for trial factoring:
https://en.wikipedia.org/wiki/GeForce_10_series

It has ~8200 GFLOPS single precision performance (FP32 which is 32 bit variables) which is good for trial factoring and general graphic performance, but "only" ~250 GFLOPS double precision (FP64, 64 bit variables) which is what is used for LL test.

Titan Black as I was comparing to has only ~5100 GFLOPS FP32 performance, but instead ~1700 GFLOPS FP64 performance, so in theory it should be 6-7 times faster at LL but slower at trial factoring, where I am only getting ~550-600 GHzdays / day.
https://en.wikipedia.org/wiki/GeForce_700_series

[storm5510]

I have an exact match. 30 hours and two minutes.

M( 46662349 )C, 0x21d7915303c760a1, offset = 3884, n = 2592K, CUDALucas v2.06beta

Originally Posted by ATH

Titan Black as I was comparing to has only ~5100 GFLOPS FP32 performance, but instead ~1700 GFLOPS FP64 performance, so in theory it should be 6-7 times faster at LL but slower at trial factoring, where I am only getting ~550-600 GHzdays / day.

It seems that Titan Black's were tailored for something different and not the obvious, gaming. In theory, as you say, the time to test this particular exponent would be much less than 30 hours. How much less, I wonder?