mfaktc: a CUDA program for Mersenne prefactoring

[kladner]

Since this falls under the hardware heading, I thought I'd drop it here, with a shout out to LaurV.

Here, and here are two pairs of EVGA water cooled 580s.

Just saying.....

[LaurV]

Originally Posted by kladner

with a shout out to LaurV. Here, and here are two pairs of EVGA water cooled 580s.
Just saying.....

There is only one pair, sold by the same guy, once piece by piece (at one link) and then both together (second link). But the same pair.
OTOH, I still have one pair of 580's full water cooled in the drawers, under the desk, for which no mobo/cpu/box is yet available (have the 1.2KW power supply, new, tho). If I buy one more mobo/box I risk that mrs LaurV moves me into the garden... I am seriously thinking to give them away, if someone can find a good home for them. By good home I mean cheap electricity, and TF or LL production for at least a while....

[Gordon]

Originally Posted by LaurV

There is only one pair, sold by the same guy, once piece by piece (at one link) and then both together (second link). But the same pair.
OTOH, I still have one pair of 580's full water cooled in the drawers, under the desk, for which no mobo/cpu/box is yet available (have the 1.2KW power supply, new, tho). If I buy one more mobo/box I risk that mrs LaurV moves me into the garden... I am seriously thinking to give them away, if someone can find a good home for them. By good home I mean cheap electricity, and TF or LL production for at least a while....

I had a 580 which died, haven't replaced it yet and when I do it won't be another 580.

580 - 244w - 433 days/day
980 - 165w - 523 days/day

I could even run two 960's for the electricity of one 580 and get an extra 80 days/day

Oh yes, forgot to mention the 580 sounds like a jet aircraft taking off and they are all quite old now...

[Xyzzy]

Originally Posted by Gordon

980 - 165w - 523 days/day

You should be able to get ~550 to 625 a day with a 980.

We have posted wattage and performance figures here: http://www.mersenneforum.org/showthread.php?t=19910

[Mark Rose]

Originally Posted by LaurV

There is only one pair, sold by the same guy, once piece by piece (at one link) and then both together (second link). But the same pair.
OTOH, I still have one pair of 580's full water cooled in the drawers, under the desk, for which no mobo/cpu/box is yet available (have the 1.2KW power supply, new, tho). If I buy one more mobo/box I risk that mrs LaurV moves me into the garden... I am seriously thinking to give them away, if someone can find a good home for them. By good home I mean cheap electricity, and TF or LL production for at least a while....

I've always wanted to try water cooling...

[TheJudger]

Compute Capability 3.7 anyone? Seems I was right, Tesla K80 is 3.7. So K80 is not Titan Z?!
K80 has two chips with 2496 cores each (13 * 192), AFAIK the chip doesn't have more, but I'm not sure. At least it shouln't be a GK110(b) like Tesla K20/K40, GTX 780 (Ti), GTX Titan (Black).

CUDA device info
  name                      Tesla K40m
  compute capability        3.5
  max threads per block     1024
  max shared memory per MP  49152 byte
  number of multiprocessors 15
  CUDA cores per MP         192
  CUDA cores - total        2880

vs.

CUDA device info
  name                      Tesla K80
  compute capability        3.7
  max threads per block     1024
  max shared memory per MP  114688 byte
  number of multiprocessors 13
  CUDA cores per MP         192
  CUDA cores - total        2496

More shared memory and (not visible in mfaktc output) double amount of registers per multiprocessor. Shared memory is 64kiB - 16kiB = 48kiB for 3.5 and 128kiB - 16kiB = 112kiB for 3.7.
So either they had those extra registers and shared memory unused all the time for other products or they really built a new chip just for Tesla K80. While chipmakers tend to built as few as possible chips this makes we wonder... AFAIK nvidia sells

• lots of Geforces
• some Quadros
• very few Teslas

(relative numbers)
And the new chip is a Kepler architecture, not Maxwell... I'm really curios!

Oliver

[tha]

I extracted mfaktc 0.20 to a directory. As far as I can tell there is no need to recompile any source. If I enter

./mfaktc.exe -h

it returns

error while loading shared libraries: libcudart.so.4: cannot open shared object file: No such file or directory

That file is in the subdirectory 'lib'. What is the advised way to proceed?

[Mark Rose]

Originally Posted by tha

I extracted mfaktc 0.20 to a directory. As far as I can tell there is no need to recompile any source. If I enter

./mfaktc.exe -h

it returns

error while loading shared libraries: libcudart.so.4: cannot open shared object file: No such file or directory

That file is in the subdirectory 'lib'. What is the advised way to proceed?

try running sudo ldconfig

if that fails, make sure that libcudart.so.4 is found in one of the directories listed in the files under /etc/ld.so.conf.d and if not add the missing directory, and then run sudo ldconfig again.

[TheJudger]

tha:

LD_LIBRARY_PATH="./lib/" ./mfaktc.exe -h

[tha]

works like a charm.

[tha]

- What does the -v(erbose) option do, I did not notice a difference between 0 and 1?
- How do I limit the output to say once per 20% progress instead of 0,1%?