David and Halsall fighting (a.k.a. Tedious bickering. Take #2)...

[Dubslow]

Originally Posted by chalsall

But don't worry too much about the "GHz Days Saved" metric. It will be going away in a while (I hope before the end of the month) -- James H. has proposed a very elegant solution (equation) which will be more representative of the value (worth) of each factor found as a function of Candidate size and Bit Level.

After all, the work saved metric was really only incentive for nucleon and Xyzzy (who were the huge guns at the time) to switch back to LL TF from DC TF.

May we please see this oft-told-of but not-actually-presented formula for "value to project of a given factor"?

[davieddy]

Originally Posted by Aramis Wyler

Yeah, maybe, I don't know. I try to run them from bottom to top, because it frustrates me when people are rolling along the bottom (like 71) and having huge marks for days saved because they find all these factors, and my factor find is the pits because I'm only doing from 73 to 74 and it takes hours to do those. I figure even the lowest end gpus capable of running mfactc can probably factor to 73 as quickly as a high end card can do 74, and that's hundreds of times faster than a cpu.

I should maybe have committed more thought to this post before I started it, but all I'm really saying is that even low end (but capable) gpus are so much better at TF than the best cpus as to be in a whole different category.
My cpus give me about 3 ghz days per core, or 12 ghz days per day. It's not the best cpu, but I like it. My video cards give 390 and 420 ghz days per day (a 480 and a 580). Even a gtx280 or a gtx450 aught to give at least 220 ghz days per day and that's so much more than 12 or 20 (for a good cpu).

It sounds to me like you are doing splendid work.

As you noticed, 70-73 takes less than twice as long as 72-73, and will find 3 times as many factors. This is why I keep harping on about doing 70-71 when the target is 73 being in a sense counterproductive in that it decreases incentive to finish the job.

Chris has noted that it is currently a bit of a struggle to ensure all new LL assignments are TFed to 73.
(See my post about "comfortable")

Also it was decided on a mixture of theoretical and practical grounds that 57M was a suitable exponent at which to TF to 73 instead of 72.
The simple 2^1/3=1.26 rule I discussed suggests that 72M is where routine TF should be upped to 74 bits.

David

[chalsall]

Originally Posted by Dubslow

May we please see this oft-told-of but not-actually-presented formula for "value to project of a given factor"?

Actually, it was-actually-presented....

[Aramis Wyler]

I see that I overestimated the lower-end cards in their throughput. I don't spend enough time on te mersenne.ca website and it seems that the 280 only puts out about a third of what I thought (67 gdays/day). Still way better than a cpu for TF, but I could totally see them being categorized together.

Originally Posted by davieddy

As you noticed, 70-73 takes less than twice as long as 72-73, and will find 3 times as many factors. This is why I keep harping on about doing 70-71 when the target is 73 being in a sense counterproductive in that it decreases incentive to finish the job.

You're right that it is much more rewarding feeling to talk the numbers bredth first. I get a little hung up on that though because 1. the reward one gets from doing it breadth first makes it even more misreable to do the top and 2. the numbers are delivered from the bottom up no matter how far they are factored once they meet the very low minimum bounds for p95 to release.

I think that maybe the 1st one wouldn't matter except for the second one, because if we go breadth first up into the 70M range then all of the numbers in the 59M range are going to get released at a bitdepth of 69. We are on a timer, and that changes the game. We're not simply trying to factor all the numbers - we're trying to get the numbers as close to well-factored as we can before they are delivered. If we had enough fire power that we could go breadth first though any given range and still have the lowest ones up to [insert appropriate bitlevel here] before they got released that would be great but 1. we don't and 2. the last couple passes would be getting the entire range up to that last bitlevel (what a grind!). That is why I think that no candidate should be let go before it is at the highest release level we can achieve short of optimal.

Originally Posted by chalsall

Yeah... However, one of the touchstones of GIMPS is that everyone should be allowed to do whatever type of work they enjoy, so long as it doesn't harm the project nor others.

That's entirely true! It doesn't matter in the slightest what I think about what people are doing, because just about the most important thing is that people can do whatever the heck they want as long as they are not hurting the project. Probably th eonly interesting debates are what the theoretical optimal factoring level is and how close can we get to that without being overtaken by the clock. Chalsall said we'd revisit that in a few weeks when primenet starts recycling the shirked assignments from the surge. No argument here.

Originally Posted by davieddy

The simple 2^1/3=1.26 rule I discussed suggests that 72M is where routine TF should be upped to 74 bits.

I can't find the anchor for your forumula there. It seems to require a constant, a determined level of propriety that the rest of the line can be drawn from, and as near as I can tell that constant is arbitrary. Maybe I missed it in a previous post.

[kracker]

Originally Posted by Aramis Wyler

...

Even a gtx280 or a gtx450 aught to give at least 220 ghz days per day and that's so much more than 12 or 20 (for a good cpu).

According to James's site on TF benchmarks here, A Geforce GTX 550 Ti produces around 130, GTX 560 does ~200.

Hmm, I wonder if nVidia or AMD is better per $... Yes I know mfakto does not have gpu sieving (yet)

[davieddy]

Originally Posted by R.D. Silverman

The last comment is amusing. My opinion is that it applies to
both of you.

Chris and I are veterans on the forum, and know each other quite well by now.
I hope I also speak for Chris when I say:
1) We don't mind the occasional ding-dong and can take it as well as dish it out.
2) Constructive things often result.

I can see why some onlookers might think we must be sworn enemies.
But...

What comparison is there between this banter and insulting complete newbies?

D

[chalsall]

Originally Posted by davieddy

What comparison is there between this banter and insulting complete newbies?

Indeed!

[davieddy]

Originally Posted by Aramis Wyler

I can't find the anchor for your forumula there. It seems to require a constant, a determined level of propriety that the rest of the line can be drawn from, and as near as I can tell that constant is arbitrary. Maybe I missed it in a previous post.

Time for a LL test = a*exponent²
Time to TF 72-73 = b/exponent.
Time to TF 73-74 = 2*time to TF 72-73
Probability of finding a factor ~1/73 in both cases.

Do you want further help?

D

[Aramis Wyler]

Nope, looks like I'm about done here. I'm tempted to plot a line in excel and post it on here to show that your forumula 'proves' we shouldn't factor them past 64, but there doesn't seem to be a point.

I'm just going to back out of this thread and regret joining it.

[davieddy]

Originally Posted by Aramis Wyler

Nope, looks like I'm about done here. I'm tempted to plot a line in excel and post it on here to show that your forumula 'proves' we shouldn't factor them past 64, but there doesn't seem to be a point.

I'm just going to back out of this thread and regret joining it.

Please don't

David

[davieddy]

Originally Posted by chalsall

Indeed!

I can hear the boos from the crowd already.
They were expecting a better fight than this

x