CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW)

[msft]

Hi, nucleon

Originally Posted by nucleon

How long are the 20M exponent double checks taking to complete?

4days.

My GTX260 share tuning and debug, thank you

[garo]

Very nice msft. That is faster than my Core2 Duo.

[frmky]

Originally Posted by msft

4days.

And a first-time LL around 50M should take just over 2 weeks.

[mdettweiler]

Originally Posted by garo

Very nice msft. That is faster than my Core2 Duo.

Same here--a while back I did a 17M doublecheck with a Core 2 Duo E4500 and it took about a week (of actual time when mprime was running). It looks like MaclucasFFTW on GPUs is really beating the socks off of standard PCs...makes me wish I had a real graphics card instead of just integrated graphics. (I don't play graphics-intensive 3D games, so I didn't see it as a priority when buying parts for my computer, though that may change now! )

Very impressive indeed, msft!

[Prime95]

Originally Posted by frmky

And a first-time LL around 50M should take just over 2 weeks.

Since only power-of-2 FFT lengths are supported, the best range for GPUs is ~35M-40M (either first-time or double-checks)

[msft]

Hi, garo

Originally Posted by garo

That is faster than my Core2 Duo.

I know your Core i5 is more fast.

Hi, mdettweiler

Originally Posted by mdettweiler

I had a real graphics card instead of just integrated graphics.

My former graphics card "S3 Trio64", is Best.

Hi, Prime95

Originally Posted by Prime95

Since only power-of-2 FFT lengths are supported, the best range for GPUs is ~35M-40M (either first-time or double-checks)

Agree,Most effective of PrimeNet's GHz days.

Thank you,

[frmky]

Another 3 of 4 verified. Again, I anticipate that the fourth is fine as well since it's a different GPU than last time.

M22033313
M22052533
M22071367
M22094129

[henryzz]

Originally Posted by frmky

Another 3 of 4 verified. Again, I anticipate that the fourth is fine as well since it's a different GPU than last time.

M22033313
M22052533
M22071367
M22094129

how error prone are gpus in comparison with cpus?
i am guessing no one knows
6/8 doublechecks matching doesnt sound good to me

i have a few questions about mlucas-cuda on 64-bit windows 7
does it write checkpoint files often so that progress can be resumed?
will it cope with hibernation? msieve-gpu wouldnt
will it use up much cpu time?

[sdbardwick]

Originally Posted by henryzz

how error prone are gpus in comparison with cpus?
i am guessing no one knows
6/8 doublechecks matching doesnt sound good to me

6 matching with 2 unverified sounds rather good to me; most likely the non-matching residues are due to mistakes in the first test. This is extrapolated (so there is an inherent uncertainty) from my own experiences when dealing with non-matching residues; if a system produces a series of matching residues, the other system is most likely at fault.
Time (and triple-checking) will tell...

[frmky]

Originally Posted by henryzz

how error prone are gpus in comparison with cpus?
i am guessing no one knows
6/8 doublechecks matching doesnt sound good to me

I'm anticipating that the non-matching ones are errors in the first test. This is not an unusual number of failed double checks. However, these are Tesla coprocessors which should be much less error prone than the typical GPU. Not sure what the error rate is on a normal consumer GPU.

Originally Posted by henryzz

does it write checkpoint files often so that progress can be resumed?
will it cope with hibernation? msieve-gpu wouldnt
will it use up much cpu time?

Yes.
If msieve-gpu doesn't, then I suspect this doesn't as well since it uses the same CUDA drivers.
Very little, probably 5-15% of one core.

[henryzz]

Originally Posted by sdbardwick

6 matching with 2 unverified sounds rather good to me; most likely the non-matching residues are due to mistakes in the first test. This is extrapolated (so there is an inherent uncertainty) from my own experiences when dealing with non-matching residues; if a system produces a series of matching residues, the other system is most likely at fault.
Time (and triple-checking) will tell...

ok i thought i remembered someone saying that the error rate used to be 4% but had risen lately due to more overclocking
i wouldnt have thought that it would have risen to 12.5%
i suppose the sample size is too small to make judgements with that information