[QUOTE=chalsall;379314]Just a heads up...

[SNIP]

And, please, let me know if anyone thinks doing this is sub-optimal.[/QUOTE]

Looks good on my end. I just grabbed a couple hundred more of whatever GPU72 wanted me to work on next and with a few exceptions (poor abandoned little things) they're all 73 to 74 in the upper 66M.

It's nice to see we're pushing everything to 74 now as it indicates to me we have a good buffer now. Also it's nice to see that the lowest assignments are in the 66M area. I didn't know we were so far ahead of the wave.

I'm still taking a mix of 68-->69 (45M range) and 69-->70 (34-35M) unless extra firepower is needed to stay ahead of the wave. My AMD GPUs are less efficient in doing 73-->74, so I'll leave that to the people with NVidia cards.

Are we even close to having the amount of firepower needed to do 75 bits? When would the ideal time be to start that?

[QUOTE=TheMawn;379561]Are we even close to having the amount of firepower needed to do 75 bits? When would the ideal time be to start that?[/QUOTE]

It doesn't really make sense to until 74M or so if I read [url=http://www.mersenne.ca/cudalucas.php?model=12]this graph[/url] correctly. Rather than trial factor to 75 bits, it makes more sense to run CUDALucas. AMD cards should only factor to 73 bits as mfakto slows down significantly beyond that from what I understand.

[QUOTE=Mark Rose;379566]It doesn't really make sense to until 74M or so if I read [url=http://www.mersenne.ca/cudalucas.php?model=12]this graph[/url] correctly. Rather than trial factor to 75 bits, it makes more sense to run CUDALucas.[/QUOTE]

Depends on the card... For 580s (GF110 chip), for example, it's "profitable" to go to 75 starting at 65M. For 680s (GK104) not until 70M. And for Titan's (GK110) not until 83M.

[QUOTE=Mark Rose;379566]AMD cards should only factor to 73 bits as mfakto slows down significantly beyond that from what I understand.[/QUOTE]

Correct.

But this is largely moot -- we don't have the fire power at the moment to sustain going to 75 (although those who want to are free to, and some are already).

[QUOTE=TheMawn;379544]I didn't know we were so far ahead of the wave.[/QUOTE]

Actually, we're not really (yet), but we're getting there. Keep in mind that the [URL="http://www.mersenne.org/thresholds/"]Primenet assignment rules[/URL] for LL Category 4 assignments is destined to be offset by 100,000 candidates (it's at 69,560 at the moment, climbing by 120 a day).

If the offset was 100,000 today the assignments would be starting at 68,562,079, and by the time it reaches the 100,000 offset it will probably be in the 70M range or so.

I think we should really think about putting in another option for which GPU is used on the PCs for optimal "Let GPUto72 decide". Some people here are using the GTX 5xx cards where it is good to go to 74 Bits from 50M, some more use the 6xx cards to go to 74 Bits from ~55M, but I wonder how many people are using the very new Cards or the AMD Cards, where they really shouldn't be doing work to 74 Bits there at all.

So I think instead of taking this work into the "Let GPUto72 decide" only into the low-exponent work or even another one, saying for which cards it is profitable.

Another suggestion would be to add a text to the get-assignments-page saying to check on the analysis on where to push the exponents, so they can adjust the limits.

The complication in that idea is when you combine multiple pieces of hardware. If the entire project were running on a single GTX 770, then yeah, by all means, TF to 73 and then LL. However, because a GTX 580 is also available, it might make more sense to TF to 74 bits first. Maybe the GTX 770 does only LL and the GTX 580 does enough TF to keep both cards fed and then LL during the rest of the time?

It gets even worse when you involve a CPU, because TF on a CPU is just bad all around. You CERTAINLY wouldn't be doing TF to 71 and then LL. Not when GPU's exist. Clearly, the optimal way is something along the lines of "GTX 580 does as much TF to 74 as it needs, and LL the rest of the time, then the GTX 770 and the CPU both doing LL 100% of the time".

It boils down to using each piece of hardware to the best of its abilities, keeping in mind the relative amounts of each that you have. For us, we have so much "CPU" that we barely keep up by putting most GPU's to TF. The GTX 770 is an interesting example actually because just yesterday I was looking at the top LL producers and the costs of the cards, and the GTX 770 is actually the sweet spot in terms of LL iterations per dollar. If people are using their GPU's for LL, I would certainly hope they're being efficient with them by picking good LL GPU's.

I don't know if Chris is still doing this but at some point he was offering to trade LL results from his CPU's for TF results from other people's GPU's in order to squeeze more efficiency out of the community. Suppose you really want to do lots of LL tests. You put all your GPU's to LL work, despite the fact that they're WAY better off doing TF. What Chris did was offer to give you the credit for his LL results done on a CPU in exchange for the credit for TF results done on your GPU's instead of LL. In the end, you get the credit for doing the LL's just like you wanted to, but overall, the project's productivity goes way up because both pieces of hardware were doing the optimal thing.


CPU's basically throw a big wrench into the optimization problem.

I think most people that have AMD cards and use GPUto72 know their cards are less efficient doing TF over 73bits. They most likely will set the 'will factor to" to 73 bit. Of course it would be interesting to see who has what kind of cards/firepower.

[QUOTE=VictordeHolland;379623]I think most people that have AMD cards and use GPUto72 know their cards are less efficient doing TF over 73bits. They most likely will set the 'will factor to" to 73 bit. Of course it would be interesting to see who has what kind of cards/firepower.[/QUOTE]

When you Let GPU72 Decide, it ignores your "will factor to" setting.

On the Work Distribution Map (primenet)
34M range Assigned TF: 69

[URL]http://en.gpu72.com/reports/available/[/URL]
34M range Reserved from Primenet 3,121 Assigned TF: 468

That doesn't seem right?