GPU to 72 status...

[James Heinrich]

Originally Posted by Mark Rose

Is there enough info in this post to add to the benchmarks? If not, perhaps TheJudger could send you what's needed.

I have 3 benchmarks (GTX 970/980) for mfaktc and they all line up pretty good, it's on CUDALucas that I only have a single benchmark.

[Mark Rose]

Originally Posted by James Heinrich

I have 3 benchmarks (GTX 970/980) for mfaktc and they all line up pretty good, it's on CUDALucas that I only have a single benchmark.

Ahh, gotcha. Perhaps you could ask TheJudger for another?

[VictordeHolland]

Originally Posted by Prime95

It looks like this card has even lower breakevens than the 570, though not by much. Some cards have TF/LL crossovers a full bit lower.

4) Make sure we don't hand out TF assignments above the crossovers in James' tables. I don't think we are near the crossovers right now.

For many cards it would be beneficial to go to 75bits for LL candidates 70M and higher and 72bits for DC candidates >40M, but we currently don't have the firepower to do so.

5) Expire TF assignments in 60(?) days. Does that sound reasonable?

30 days with the possibility to extent them has worked great for GPU72 in the past, but I don't see any problem with 60 days.

7) Have GPU72 return the assignments it hasn't handed out. Have GPU72 forward TF requests to PrimeNet.

It would be nice to see the Work Distribution page provided the 'complete picture' again. Now almost the entire 70-79M range is reserved by GPU72, while only 3,000 out of the 167,000 exponents are actively worked on.

[Mark Rose]

Originally Posted by kladner

Thirty days should be more than enough. At least, that is the current GPU 72 period. I personally start checking what the status of a factoring job is if it gets to be 8-10 days old. I expect them to be gone before then. I know some people keep more work in the hopper than I do. Perhaps a longer time allowance would suit them better.

I agree. Even on an old, slow card, such as a GT 520 or GT 430, a 70->75 75M LLTF assignment could be completed in 2 to 4 days. With an automated system even two weeks is more than enough. Might be a nuisance for people who do things manually though.

[TheMawn]

Would it be possible to assign DC's that have not been appropriately TF'ed (say, one bit level short of optimal) to people who have never once reported a result?

Every single person would do at most one less-than-optimally-TFed job in their entire life but it would would divert all properly factored exponents to the people way more likely to complete them.

[NickOfTime]

Originally Posted by TheMawn

Would it be possible to assign DC's that have not been appropriately TF'ed (say, one bit level short of optimal) to people who have never once reported a result?

Every single person would do at most one less-than-optimally-TFed job in their entire life but it would would divert all properly factored exponents to the people way more likely to complete them.

Well, we shouldn't need to, we are ahead of the cat 4 churn now 1800 vs 1400 assigned. http://www.gpu72.com/graphs/dctf/week/

[Prime95]

I'm working on the Primenet GPU TF assignment page. It works much like the GPU72 page.

Two questions:

1) The system is designed to work one bit-level at a time. Will this create some work units that are just too short? Should I modify the assignments to do multiple bit levels if the current bit level is somewhat low (or should I leave this under user control by filling out the "optional will factor to" field)? If automatic, what are the recommended minimum bit level to factor to for a DC, LL, and 100M?
2) What is the current TF/LL crossovers for GPUs on 100M digit numbers?

[Mark Rose]

Factoring to bit levels at or below 69 should probably be assigned as a single assignment. At 80M, that's about 1.5 GHz-days, and even an ancient card can do several of those per day. A card like a GTX 580 or GTX 970 can do about 300 of those a day.

I would let users pick a maximum bit-level, minimum 69. If we start factoring to 75, a 71->75 at 75M would take about 95 GHz-days. It could discourage new users if they have slow cards that can't finish a single assignment in a day. I think doing the same thing with LL/DC categories and assign high work to users who have never returned TF work might be a good idea. GPU72 current limits the amount of work given out to new users and PrimeNet should do the same.

Until recently, AMD cards had a penalty factoring beyond 73 bits (I don't know if the version of mfakto has been released with the newer kernel). It might be useful to add a suggested bit-level field (hidden?) for automated clients to pass that indicates at what bit level performance drops. Those clients could still work on higher bit levels if that's what the system needs, but ideally would work on lower bit levels for overall system throughput. That's one feature "Let GPU72 decide" lacks.

According to mersenne.ca, the 100M cross-over level for a GTX 970 is 77 bits LLTF and 76 bits DCTF. For a GTX 780 is 76 and 75. I doubt that takes into account the severe performance hit mfaktc has above 76 bits. This is where that suggested bit-level field would come in handy.

[VictordeHolland]

Originally Posted by Mark Rose

According to mersenne.ca, the 100M cross-over level for a GTX 970 is 77 bits LLTF and 76 bits DCTF. For a GTX 780 is 76 and 75. I doubt that takes into account the severe performance hit mfaktc has above 76 bits. This is where that suggested bit-level field would come in handy.

I think you might be mixing up 100M 'exponent' (2^100,000,000-1) and 100M digits (2^332,200,000-1).
For 100M exponent the cross-over is indeed +/- 77bits.
For 100M digits candidates the 'normal/CPU' TF bitlevel is already 77bits, so with GPUs you could probably do 4-5bits more, so 81bits, maybe even 82bits if there are enough resources.

[Prime95]

Originally Posted by Mark Rose

According to mersenne.ca, the 100M cross-over level for a GTX 970 is 77 bits LLTF and 76 bits DCTF. For a GTX 780 is 76 and 75. I doubt that takes into account the severe performance hit mfaktc has above 76 bits. This is where that suggested bit-level field would come in handy.

I'm asking about 100M digits or M332192000. I think the chart you refer to is for M100000000.

A prototype of the PrimeNet web page is: http://mersenne.org/manual_gpu_assignment/
Feel free to click on getting assignments, it will display work without making any real reservations. Note that the assignments returned is not what you'll eventually get since GPU72 has nearly all the relevant DC and LL exponents reserved.

[James Heinrich]

Originally Posted by Mark Rose

According to mersenne.ca ... I doubt that takes into account the severe performance hit mfaktc has above 76 bits.

Correct. My performance charts are based on a simple 1-dimensional measurement, it does not scale appropriately non-linearly where different bit levels or kernels are invoked. Something I should probably look at in the future, I guess.