[QUOTE=Madpoo;475511]I sure thought that'd be a cool use of my $200 credit on my personal account but oh well. Maybe next time.[/QUOTE]

Yeah. You have to beg and plead to get access to the latest instances under Azure.

I did, and I found they were *much* more expensive than Amazon's spot instances, or even Google Instances.

[QUOTE=Madpoo;475511]The NC6s_v3 is currently showing up at $1.53/hr and if it took something like 30 hours (am I even close? No idea...) then we're talking $46 USD for a rapid verification run.[/QUOTE]Some of us would be willing to toss a few quid into the kitty to have a reserve to pay for this. No reason that the forumites can't pay for some time on some hot hardware.

I wonder if Alphabet or Bezos or Gates would be willing to donate time on top notch machines for verification runs. Might be worth it to them for the press. Maybe something that can be set-up for next time.

[QUOTE=Prime95;474943]A day after Christmas a computer has reported a new Mersenne prime. Unlike the numerous CudaLucas false positives we've had over the last few years, this one shows promise. It was assigned December 18, no errors during the run. This particular computer has reported 57 previous results including some successful double-checks.

Madpoo's running an LL test now. T-41 hours.

I'll reach out by email for the last save file, but with the holidays I may not get a response before Madpoo's run completes.

Fingers crossed!![/QUOTE]

I hope this works out! :smile:

[QUOTE=storm5510;475530]I hope this works out! :smile:[/QUOTE]
We all do.. wait... did. What is this, a time machine? We are in the 26th of December again!?

Phew. No, it is the 30th, and we have a definitive answer.

What a bad dream!

[QUOTE=Madpoo;475511]I'd looked at doing something similar with an Azure GPU instance (Tesla graphics cards) but they didn't show up in my current subscription. The NCv3 instances with a Tesla v100 are apparently only available on request.

The NC6s_v3 is currently showing up at $1.53/hr and if it took something like 30 hours (am I even close? No idea...) then we're talking $46 USD for a rapid verification run.

I sure thought that'd be a cool use of my $200 credit on my personal account but oh well. Maybe next time.[/QUOTE]

I spun up a V100 on my AWS account about midway through my Vega run. Timing was fast (I think I sent Ernst the numbers but don’t have them offhand), but not fast enough to beat the Vega’s already smoking run. I’m glad I tried it though, convinced me not to buy a Titan V on my EOY R&D budget for “testing”

[QUOTE=airsquirrels;475533]I spun up a V100 on my AWS account about midway through my Vega run. Timing was fast (I think I sent Ernst the numbers but don’t have them offhand), but not fast enough to beat the Vega’s already smoking run. I’m glad I tried it though, convinced me not to buy a Titan V on my EOY R&D budget for “testing”[/QUOTE]

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?

Re. my previous post about c5.9xlarge timings, I should clarify - the 2.4 msec/iter @4096K I got was @16 threads, more threads no help. Similar for ATH's 3.0 ms/iter @4608K, which best-time was @18 threads, again any more hurt rather than helped. Not being a regular AWS user, can someone tell me - the c5.9large has 36 (logical) cores, would running 2 such 16-or-18-threaded jobs side-by-side to more fully utilize said cores cost any more than just one job?

[QUOTE=airsquirrels;475533]I spun up a V100 on my AWS account about midway through my Vega run. Timing was fast (I think I sent Ernst the numbers but don’t have them offhand), but not fast enough to beat the Vega’s already smoking run. I’m glad I tried it though, convinced me not to buy a Titan V on my EOY R&D budget for “testing”[/QUOTE]

This png which you sent me, right?

[QUOTE=Uncwilly;475521]Some of us would be willing to toss a few quid into the kitty to have a reserve to pay for this. No reason that the forumites can't pay for some time on some hot hardware.

I wonder if Alphabet or Bezos or Gates would be willing to donate time on top notch machines for verification runs. Might be worth it to them for the press. Maybe something that can be set-up for next time.[/QUOTE]

Somehow I can't picture multi-hundred-billion-dollar corporations issuing press releases touting the fact that they donated... the equivalent of a modest restaurant dinner. No PR benefit there, it would just look silly.

Now, if by any chance they wanted to donate a $10K cloud credit for general Mersenne number crunching...

Btw the corresponding Perfect number +1 has small factors:
2[sup]p-1[/sup]*(2[sup]p[/sup]-1)+1

and the corresponding Wagstaff number has a factor (NMC):
2[sup]p+1[/sup] / 3

Still no factor of 2[sup]p[/sup] - 3 (twin prime)

[QUOTE=ewmayer;475535]Re. my previous post about c5.9xlarge timings, I should clarify - the 2.4 msec/iter @4096K I got was @16 threads, more threads no help. Similar for ATH's 3.0 ms/iter @4608K, which best-time was @18 threads, again any more hurt rather than helped. Not being a regular AWS user, can someone tell me - the c5.9large has 36 (logical) cores, would running 2 such 16-or-18-threaded jobs side-by-side to more fully utilize said cores cost any more than just one job?[/QUOTE]

I won't share my per-iteration time since I already mentioned a total runtime of around 36 hours for me (Prime95 v29.4b5) and the iteration time would be a near dead giveaway.

Beyond that, it's a dual 14-core Xeon and typically I run 2 workers of 14 cores each, using all the cores on each socket. Works well.

When I started that way, it gave me an ETA of 42 hours, I think, and then I reconfigured to have a single worker with 20 cores (14 on one socket, 6 from the other).

That ended up giving me the best speed overall. With more than 20, the timings actually got worse, which is a phenomenon I'd seen in previous testing. The QPI link only has X amount of bandwidth and I think it gets overwhelmed when too many cores are sending their bits to the main thread.

Anyway, the AWS and Azure VMs with a lot of cores are almost definitely multi-socketed systems. Plus, newer Azure VMs use hyperthreading which means that shiny multi-core VM might not be all that awesome for FP64.

There may or may not be accurate NUMA mapping from the hardware side to the VM. Hopefully there is (Hyper-V seems to show it okay, but I've never really got into it). That may mean you'd have to check a few things out to see just how the hypervisor assigned physical cores to your particular box.

My verify run @4M FFT length just finished - no surprise, it's prime. Congrats to all!