Prime95 version 28.9 / 28.10

[aurashift]

Originally Posted by Mark Rose

Have you tried benchmarking with BenchMultithreads=1 in prime.txt to see what the actual throughput would be?

seems to be doing the same thing...although mprime is doing something consistently regardless of this setting when doing timings. When I get past 40 threads it'll just freeze up.

[rudi_m]

For me 28.10 also crashs when having TF workers. It happens always on all of my machines.

Since this seems to be a known problem it's a bit annoying that users still have to run into this issue, IMO this version should be removed from the official download page.

[pepi37]

Currently I do make test on 500K digit PRP. I divide test range on 2 CPU , one is Ivy Bridge , second is Haswell, and I give 30% more work to Haswell because it has FMA3
What I noticed, even with 30% more work , for now , there was none warning on Haswell but there was ( in same time) 7 warning on Ivy Bridge considering
Confidence in final result is fair.
Disregard last error. Result is reproducible and thus not a hardware problem.
For added safety, redoing iteration using a slower, more reliable method.

So it look like that on same OS ( Linux Debian), same version mprime28.10 FMA3 is much more resistant to iteration problem.

That is just my notice, observation

[debrouxl]

Like aurashift above, I get consistent SIGILL with 28.10 when doing TF, on a Skylake (Core i7-6700QM) computer, even after disabling AVX + FMA3. libhwloc is already installed on the computer.
Pretty please, George ?
TIA.

[Prime95]

For TF users only:

Link to Windows 29.1: https://www.dropbox.com/s/5vjhsu9gbo...win64.zip?dl=0

Link to Linux 29.1: https://www.dropbox.com/s/a53l99b68u...ime64.tgz?dl=0

TF code is still being worked on, so this version will be replaced soon.

[Uncwilly]

Originally Posted by Prime95

For TF users only:

I see an ~8% speed boost for work on exponents in the 332,000,000 range, vs. v28,.9.
The 2 versions are reporting vastly different completion predictions for the range (and neither correct for the actual completion). v28.9 is reporting ~6 days for such exponents going from 75 to 76, v29.1 is reporting 12 hrs. Actual estimated time (based upon the wall clock) should be 26 hours (for v29.1).

[Gordon]

Originally Posted by Uncwilly

I see an ~8% speed boost for work on exponents in the 332,000,000 range, vs. v28,.9.
The 2 versions are reporting vastly different completion predictions for the range (and neither correct for the actual completion). v28.9 is reporting ~6 days for such exponents going from 75 to 76, v29.1 is reporting 12 hrs. Actual estimated time (based upon the wall clock) should be 26 hours (for v29.1).

Just for interest I put mfaktc to work on 33230002 from 75-76 bits

...about 82 mins, or about a factor of 20 faster on a GTX-970

[Uncwilly]

Originally Posted by Gordon

Just for interest I put mfaktc to work on 33230002 from 75-76 bits

To abuse a quote: "I trial factor with the equipment I have, not the equipment I wish I had."


Also, I went to put 29.1 on another machine and found out that there is a .dll that is needed.

[VBCurtis]

Originally Posted by Gordon

Just for interest I put mfaktc to work on 33230002 from 75-76 bits

...about 82 mins, or about a factor of 20 faster on a GTX-970

You know time varies by exponent, right? You tested a 33M exponent, while the person you quoted tested 332M.

[Gordon]

Originally Posted by VBCurtis

You know time varies by exponent, right? You tested a 33M exponent, while the person you quoted tested 332M.

Didn't quite cut & past enough

UID: nitro/GTX-970, no factor for M332300027 from 2^75 to 2^76 [mfaktc 0.21 barrett76_mul32_gs]

[Prime95]

Originally Posted by Uncwilly

Also, I went to put 29.1 on another machine and found out that there is a .dll that is needed.

Get your DLLs here: https://www.open-mpi.org/software/hwloc/v1.11/