GPU TF calculations on CPU lost work with AMD GPUs

[bcp19]

Since this mainly affects GPU72 I figured I'd make a new topic here.

I did some thorough testing with one of my quads and an HD 5770 yesterday to determine the difference made when using a GPU that does not currently have LL capability. The one data set I had from when it was in a very fast machine showed it could have done 127 TFs from ^68-^69 at the 26M level in the time it took the core running it to produce a DC. I estimated putting this card in my Core 2 Duo would bump this up to 182 TFs/DC. I put the card in my Q8200 quad, and after testing all variants, 2 cores P95 and 2 instances mfakto gave the best results. In the time these 2 cores could do a 26M DC, the card would produce 286 TF from ^68-^69.

If possible, I'd like to get some information from people running mfakto. If you could PM me: CPU and GPU model (Q8200/5770), # cores running P95/mfakto (1/3, 2/2, 5/1), exp size and iter time on P95 cores (45381221/.059ms, 26202373/.035ms), exponent size, bit level and rough avg time for each mfakto instance (30311929/68-69/74m29s, 30363997/68-69/73m52s, 29499839/69-70/162m11s) it would be a great help. Thanks in advance for your help.

[KyleAskine]

I only run mfakto and p-1 on my pc's. Do you still want the mfakto info?

[bcp19]

Originally Posted by KyleAskine

I only run mfakto and p-1 on my pc's. Do you still want the mfakto info?

Yes, please, unless you happen to know your daily GD output per GPU(I notice you average around 480-500 on GPU72, is that 1 or 2 cards?). Also, if you happen to know any rough exp iter timings for DC or LL that would also be helpful, or tell me the CPU model and I can look it up on James' site for an estimate.

[KyleAskine]

Originally Posted by bcp19

Yes, please, unless you happen to know your daily GD output per GPU(I notice you average around 480-500 on GPU72, is that 1 or 2 cards?). Also, if you happen to know any rough exp iter timings for DC or LL that would also be helpful, or tell me the CPU model and I can look it up on James' site for an estimate.

Four cards, actually. Us AMD folk factor slower.

One system:
i-2500k overclocked to 4.6 GHz
2x6950 with a 6970 bios (shaders unlocked)
Just finished M59981609 from 70 to 71 with Load at 90% in 45m 42s

Another system:
AMD Phenom II 955 not overclocked (it is old and locks up if I try)
Two instances of mfakto, so I can't give you a real clear view into the timings.

A third system, that I do not have access to right this second:
A core2duo with a 6570 - it's slow.

[bcp19]

Originally Posted by KyleAskine

Four cards, actually. Us AMD folk factor slower.

One system:
i-2500k overclocked to 4.6 GHz
2x6950 with a 6970 bios (shaders unlocked)
Just finished M59981609 from 70 to 71 with Load at 90% in 45m 42s

Another system:
AMD Phenom II 955 not overclocked (it is old and locks up if I try)
Two instances of mfakto, so I can't give you a real clear view into the timings.

A third system, that I do not have access to right this second:
A core2duo with a 6570 - it's slow.

Slower, yes, but more suited. I was able to work up each system (the Phenom still has the 5870?). James' site as always was invaluable. While this is a relatively small sample, the data does show that without the GPU-LL ability, the AMD/ATI GPU/CPU combinations can easily do 0-3 bit levels more than the best nVidia GPU/CPU ones that can run CUDALucas. My nVidia graph had had enough green (well above breakeven) to fill ~40% of the grid I used. The AMD graph was almost 84% green. If you remove the worst system (mine), seeing as it was an underpowered GPU in an overpowered CPU, it would be almost 95% green.

[KyleAskine]

Originally Posted by bcp19

Slower, yes, but more suited. I was able to work up each system (the Phenom still has the 5870?). James' site as always was invaluable. While this is a relatively small sample, the data does show that without the GPU-LL ability, the AMD/ATI GPU/CPU combinations can easily do 0-3 bit levels more than the best nVidia GPU/CPU ones that can run CUDALucas. My nVidia graph had had enough green (well above breakeven) to fill ~40% of the grid I used. The AMD graph was almost 84% green. If you remove the worst system (mine), seeing as it was an underpowered GPU in an overpowered CPU, it would be almost 95% green.

You got it right, my AMD platform has the 5870.

Cool information, thanks for compiling that.