PrimeGrid
Please visit donation page to help the project cover running costs for this month

Sponsored by:

Hosted at and sponsored by Rackspace.

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
1) Message boards : Number crunching : 2018 Tour de Primes (Message 114940)
Posted 10 days ago by Iain BethuneProject donor

I didn't realize power supplies were also hard to get now! Hopefully the cryptocurrency market will take its inevitable downturn and prices will return to normal.


Interestingly, the squeeze on GPU availability due to the popularity of mining made the front page of the BBC earlier today!

- Iain
2) Message boards : Aggie The Pew message board (Message 114122)
Posted 22 days ago by Iain BethuneProject donor
this team has but one rule, and it is that there are no rules.


Isn't that rule 2?

P.S. Welcome to x/0

3) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 114117)
Posted 22 days ago by Iain BethuneProject donor
Update on OCL5 "bug" seen on 52186^131072+1:

As far as we know at present this is likely to be a hardware-related issue rather than a code bug. Round-off errors have been seen on Van's Mac with Radeon RX 480 and composite's Linux box with GeForce GTX 760. On Van's machine the error is reproducible on most runs, but not always at the same point in the calculation. On composite's it has only been seen once. Van also sees the same behaviour with the current stock app (3.3.2).

We have over 40 successful tests in both Linux/Nvidia and Mac/AMD setups, and have additional successful tests on Windows/Nvidia and Mac/Nvidia.

composite - would you be able to re-run the test several times on your host where you first saw the bug and see if appears to be recreatable?

In any case, I think we can continue with the test programme - if anyone has Linux and/or Windows systems with AMD GPUs, that could run some tests that would be very helpful - PM me if you're unsure what to do!

- Iain

4) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 113942)
Posted 25 days ago by Iain BethuneProject donor
G48-G56:

OCL5 Returned an error, and it looks like it isn't the first time it has occurred. I attempted to re-run it, and got the same result.


Would be good if you could try running this test several times - and also with the previous genefer release (current stock PrimeGrid app will do). These does seem to be an issue here that we need to understand!
5) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 113940)
Posted 25 days ago by Iain BethuneProject donor
Bit of a toaster of a datapoint compared to most things, but on my i7-4710MQ (2.5GHz) laptop (w7x64),


These results confirm pretty much what everyone else has seen - a decent speedup for the x87 transform, and roughly the same (or slightly slower, depending on hardware and n) performance for the vector transforms. Note that the b values used for the "Generalized Fermat Prime Search benchmarks" are increased in 3.3.3 so you can't directly compare. The b limit for the vector transforms are slightly increased - which might bring some tests back into range.

- Iain
6) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 113939)
Posted 25 days ago by Iain BethuneProject donor
Please note there are new Windows CPU builds (3.3.3-3), which fix a minor issue with priority when running standalone (i.e. BOINC is not affected). Anyone who is running the CPU code on Windows - please go back to https://app.assembla.com/spaces/genefer/subversion/source/HEAD/trunk/bin/windows and grab the updated binary!

Thanks for all the testing so far!

Cheers

- Iain
7) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 113924)
Posted 26 days ago by Iain BethuneProject donor
geneferocl_linux64 starts 8 threads, but only the main one gets essentially all the CPU time (another one amassed a total of 0.10 seconds by the time the main thread had 10.5 minutes, and a third got 0.01 seconds). Is this normal?


No idea - but I'm not surprised it spawns some additional threads (presumably in the OCL/driver stack, rather than the user code).

- Iain
8) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 113923)
Posted 26 days ago by Iain BethuneProject donor

geneferocl 3.3.3-2 (Linux/OpenCL/64-bit)

Copyright 2001-2018, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2014, Michael Goetz, Ronald Schneider
Copyright 2011-2018, Iain Bethune
Genefer is free source code, under the MIT license.

Command line: ./geneferocl_linux64 -q 52186^131072+1 -x OCL5

Normal priority change succeeded.

Testing 52186^131072+1...
Using OCL5 transform
Running on platform 'NVIDIA CUDA', device 'GeForce GTX 760', vendor 'NVIDIA Corporation', version 'OpenCL 1.2 CUDA' and driver '384.111'.
6 computeUnits @ 1137MHz, memSize=1996MB, cacheSize=96kB, cacheLineSize=128B, localMemSize=48kB, maxWorkGroupSize=1024.
Starting initialization...
Initialization complete (0.113 seconds).
Testing 52186^131072+1... 2054077 steps to go
Estimated time for 52186^131072+1 is 0:09:34
maxErr exceeded for 52186^131072+1, 1.0000 > 0.4500
Errors occurred for all available transform implementations

reran OCL5, it picked up from the last checkpoint
geneferocl 3.3.3-2 (Linux/OpenCL/64-bit)

Copyright 2001-2018, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2014, Michael Goetz, Ronald Schneider
Copyright 2011-2018, Iain Bethune
Genefer is free source code, under the MIT license.

Command line: ./geneferocl_linux64 -q 52186^131072+1 -x OCL5

Normal priority change succeeded.

Testing 52186^131072+1...
Using OCL5 transform
Running on platform 'NVIDIA CUDA', device 'GeForce GTX 760', vendor 'NVIDIA Corporation', version 'OpenCL 1.2 CUDA' and driver '384.111'.
6 computeUnits @ 1137MHz, memSize=1996MB, cacheSize=96kB, cacheLineSize=128B, localMemSize=48kB, maxWorkGroupSize=1024.
Resuming 52186^131072+1 from a checkpoint (2054077 iterations left)
Estimated time for 52186^131072+1 is 0:09:30
52186^131072+1 is composite. (RES=1b196d6c0e4d778f) (618340 digits) (err = 0.0000) (time = 0:09:28) 21:47:16


This is strange. b=52186 is well below the b limit, and for OCL5 the calculation should be exact (err=0.0000). Does this recreate on your hardware? It seems to run fine for Grebuloner and Van Zimmerman, and also I tested on my own linux system with a Tesla K40m. If not I can only think it is either a hardware glitch, or a (very rare) software bug. In either case the error correction worked and it restarted from the last known good checkpoint...
9) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 113921)
Posted 26 days ago by Iain BethuneProject donor
B10 reserved
...
EDIT: This was completed before the spreadsheet changed. I will run with the other transforms now



Sorry - I have updated the spreadsheet now. Please check posts in the thread to see if any tests are outstanding that are not yet in the spreadsheet.

Cheers

- Iain
10) Message boards : Generalized Fermat Prime Search : Genefer 3.3.3 testing (Message 113920)
Posted 26 days ago by Iain BethuneProject donor
C10; did not switch to X87 transform


That's fine - the b limits of the 'fast' transforms are increased a little.


Next 10 posts
[Return to PrimeGrid main page]
Copyright © 2005 - 2018 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 1.55, 1.85, 1.97
Generated 24 Feb 2018 | 15:51:49 UTC