[QUOTE=axn;475534]Did you use latest CUDA SDK (9.1) to compile CuLu for the appropriate CC (7.0)? V100 has 6 TFLOPS of DP. I don't think anything could beat that. What was the timings for Vega? There is some timings from V100 (given by TheJudger) that I have quoted in this thread -- how does you timings (both Vega and V100) compare to that?[/QUOTE]

V100 (virtualized) 1.11 ms/iteration - 9.1 CUDA latest build of cudaLucas Compiled 70.

CUDALucas picked a 4320 FFT, although we have now seen that a 4096K would have worked and been a bit faster.

Vega using gpuOwl’s 4096K FFT was 1.55ms/iteration. I wasn’t saying that the V100 wasn’t faster, just that it wasn’t fast enough to catch the Vega that already had a 10 hour head start. It would appear with the lower memory bandwidth that the Titan V would be on par or just slightly faster than than Vega for its LL performance. At 10x the price.

1.11 ms/iteration for 4320K FFT seems a bit slow for a V100. I see a stable ~0.94 ms/iteration for 4608K FFT (M81xxxxxx). Had you been able to check temperature and power consumption during the run (nvidia-smi -a is a nice tool!)

Ofc Vega is wins the performance/price comparison with a huge margin anyway.

Oliver

As long as people have CPU time to spare - they could run the prime95/gpuOWL/cudaLucas/Mlucas with
[COLOR=Blue]InitialLLValue=10 [/COLOR] (this can be obviously trivially coded into every alternative program, too)[COLOR=Blue]
[/COLOR]and[COLOR=Blue]
[/COLOR][COLOR=Blue][COLOR=Blue]InitialLLValue=23
[/COLOR][/COLOR](not of the sake of the last values of the residue, but to test how the errors behave etc).

[QUOTE=ET_;475172]If so, I'm asking for one... :razz:[/QUOTE]

To add to my collection :-)

[QUOTE=Prime95;475190]:thumbs-up::thumbs-up:People need to read that and let it sink in.

If we revert to the historical norm, we should expect 5 or 6 new primes between exponents 80,000,000 and 800,000,000. That's 5 or 6 new primes over the next 3 (or more) decades!![/QUOTE]

Except not really though is it?

From powers of 10

0-1 is just 10 candidates
1-2 is 90 more candidates
2-3 is 900 more candidates
3-4 is 9900 more candidates

Given that each range is *ten* times the size of the previous one then correcting for relative size we are missing a shed load....

[QUOTE=ATH;475537]Still no factor of 2[sup]p[/sup] - 3 (twin prime)[/QUOTE]
Simply running PRP will take less than 46 hours, Andreas! If you still don't have a factor - run PRP.

[QUOTE=Gordon;475602]To add to my collection :-)[/QUOTE]

If this new prime was not found by Central Missouri (who issued the last several buttons), we may need to take up a collection to have some commemorative buttons made. And have the forum gerbils judge a prime-button-design contest?

Norm

Just a minor "typo". May I suggest that the title of the announcement on the homepage be changed from:
"Breaking News! Computer reports discovery of 50th Known Mersenne Prime!!"
to:
"Breaking News! Computer reports discovery of 50th known Mersenne Prime!!"

The "Known" isn't a proper noun nor is the rest of the title's main words capitalized so it should be lowercase.

[QUOTE=Gordon;475604]Given that each range is *ten* times the size of the previous one then correcting for relative size we are missing a shed load....[/QUOTE]

But take a look at [URL="https://primes.utm.edu/mersenne/heuristic.html"]the graph[/URL] here, there are good reasons to believe it's a straight line.

[QUOTE=daxmick;475628]...
The "Known" isn't a proper noun nor is the rest of the title's main words capitalized so it should be lowercase.[/QUOTE]

Fixed as suggested.

[QUOTE=ewmayer;475504]FYI, Andreas (ATH) kindly gave me access to the AWS c5.9xlarge instance on which he completed his Mlucas run @4608K - he started on a mix of C4 and C5.18xlarge (and started with the slower avx2 build due to compile problems with the Mlucas 17.1 release in avx512 mode), and only late in the verify found the c5.9xlarge gave him slightly better speeds than the c5.18xlarge at half the price. I wanted to test his avx-512 build @4096K, as well - long story short, I got a best time of 2.4 msec/iter at that smaller FFT length on the c5.9xlarge, i.e. his verify would have needed just a little over 48 hours had it used that setup for the whole run. Looks like it's time for me to set up a user account for AWS. :)[/QUOTE]

[url]https://aws.amazon.com/grants/[/url]

Including for "Perform proof of concept or benchmark tests evaluating the efficacy of moving research workloads or open data sets to the cloud.", which is exactly what you're doing.