[QUOTE=Madpoo;519161]Milestone page updated.

I really should have it show something like "complete" for that milestone, until I can actually update the page. I was sure it was happening today, but of course it happened during the workday when I wasn't able to update it. :smile:[/QUOTE]Automate it. Create the page dynamically. Make the computer do the work for you.

No milestone for the first 200M digits number verified? :rant:
hehe

[QUOTE=LaurV;519238]No milestone for the first 200M digits number verified? :rant:
hehe[/QUOTE]Powers of ten only.

[QUOTE=Uncwilly;519241]Powers of ten only.[/QUOTE]So Laurv, feel free to verify the top exponent in the mersenne.org range 2 - 10[SUP]9[/SUP], 999,999,937.
We could consider powers of two as lesser milestones, but the EFF prizes are for powers of ten.

[QUOTE=kriesel;519252]So Laurv, feel free to verify the top exponent in the mersenne.org range 2 - 10[SUP]9[/SUP], [B]999,999,937[/B].
We could consider powers of two as lesser milestones, but the EFF prizes are for powers of ten.[/QUOTE]

Given the fastest known hardware that can run software to verify this number (whether it is Cray's, 128 core beasts, etc), how long would a verification take, approximately?

[QUOTE=PhilF;519274]Given the fastest known hardware that can run software to verify this number (whether it is Cray's, 128 core beasts, etc), how long would a verification take, approximately?[/QUOTE]What is currently the fastest setup, and what is the required time, are good questions. An upper bound is half a year, set by ramgeis for the first-test with an NVIDIA GV100 running CUDALucas.
[url]https://www.mersenneforum.org/showpost.php?p=495854&postcount=8[/url]
As a double check, Laurv could take ramgeis up on the offer of interim save files and do a verification using more than one GPU in parallel. So with n of the same gpu, and n-1 equally spaced interim save files, he could do it in as little as 1/n times 6 months. The electric bill might be a shocker.

[QUOTE=PhilF;519274]Given the fastest known hardware that can run software to verify this number (whether it is Cray's, 128 core beasts, etc), how long would a verification take, approximately?[/QUOTE]

Here are benchmark I did on a Tesla V100 on the Amazon EC2 instance:
[url]https://www.mersenneforum.org/showpost.php?p=493539&postcount=2684[/url]

The minimum FFT size for the ~1B exponent is 57344K and the fastest benchmark times above that is 59049K FFT at 11.42 ms/iteration.

Lets assume 12 ms/iteration then it would take 12B ms = 12M seconds ~ 139 days ~ 4.6 months.
But a Tesla V100 Amazon instance costs ~$1 per hour when they are cheapest, so 139 days would be $3,336. Compared to buying a Tesla V100 card: ~$10,000 plus the computer to use it in.

If we're to believe the numbers [URL="https://www.mersenne.ca/cudalucas.php"]here[/URL], a Radeon VII will get you there quicker, cheaper.

[QUOTE=axn;519284]If we're to believe the numbers [URL="https://www.mersenne.ca/cudalucas.php"]here[/URL], a Radeon VII will get you there quicker, cheaper.[/QUOTE]Radeon VII looks like a great choice for PRP OpenCL based first tests. But the first test for M999999937 was an LL test done in CUDALucas, and the Radeon VII won't run that. It would be necessary to use ClLucas, which takes roughly twice as long as GpuOwl for the same exponent and hardware.

[QUOTE=ATH;519283]Here are benchmark I did on a Tesla V100 on the Amazon EC2 instance:
[URL]https://www.mersenneforum.org/showpost.php?p=493539&postcount=2684[/URL]

The minimum FFT size for the ~1B exponent is 57344K and the fastest benchmark times above that is 59049K FFT at 11.42 ms/iteration.

Lets assume 12 ms/iteration then it would take 12B ms = 12M seconds ~ 139 days ~ 4.6 months.
But a Tesla V100 Amazon instance costs ~$1 per hour when they are cheapest, so 139 days would be $3,336. Compared to buying a Tesla V100 card: ~$10,000 plus the computer to use it in.[/QUOTE]
Or buy a Titan V at $3000, accept the 10-20% longer run ~170 days, and the electrical cost ~$125 (computed at US$0.12/kwhr), plus maybe $40 for air conditioning, use it as a replacement gpu in an existing system. It makes taking on the next exponent down, 999,999,751, more economical.

[QUOTE=kriesel;519302]Radeon VII looks like a great choice for PRP OpenCL based first tests. But the first test for M999999937 was an LL test done in CUDALucas, and the Radeon VII won't run that. It would be necessary to use ClLucas, which takes roughly twice as long as GpuOwl for the same exponent and hardware.[/QUOTE]

True, but it is a weekend's work for preda to bring back LL to gpuowl, if there really is a need for it.