forums.silverfrost.com

DanRRight · Posted: Sun Jun 21, 2009 6:05 am Post subject: Anyone tried NVIDIA CUDA ?

or similar parallel tools from AMD/ATI together with FTN95 to make you 100-300 core supercomputer for numbercrunching out of old home basement PC ?

DanRRight · Posted: Sat Jun 26, 2010 3:06 am Post subject:

Look what folks do with them

JohnCampbell · Joined: 16 Feb 2006 Posts: 2554 Location: Sydney

Dan,

Can you give us an overview of this ?
What is PMK ? I googled and got a Tamil Political Party.
Is there a DLL interface that will allow us to use the NVIDIA CUDA instruction set ?
It would be good to know what types of instructions are available for real*8. I am more interested in dot product and large vector operations; such as A = A - const * B, where A and B are vectors of, say 1 to 100,000 long ( as apposed to 3 or 4 for my graphical calculations).

John

Sebastian · Joined: 20 Feb 2008 Posts: 177

PMK==Pairwise Master Keys, so they are measuring the amount of generated keys per second (see http://code.google.com/p/pyrit/ ).

DanRRight · Posted: Fri Jul 02, 2010 7:55 pm Post subject:

I've never tried computing on GPU, i've done so far parallelization on multi-core CPUs via third-party libraries (LAIPE, linear algebra) and tried to use MTASK library which multitasks any job you want OK on FTN77/95 (but conflicts if you make Print* in each thread simultaneously). These libraries were built for different Fortran compilers but specifically ancient Microsoft's ones worked OK with FTN77/95. Attempts of making these libraries native by compiling them with FTN95 failed by unknown reasons. Though this is kind of pity fact, that became not a big deal lately when all compilers started supporting DLLs. If anyone here is interested and owns Intel Vis.Fortran these libraries (LIB files) can be put into DLL wrapper which most probably will be highly compatible with FTN95 and any other compiler and also will be incredibly fast since they were just recompiled on latest IVF.

BTW, how to do parallelization on CPUs easiest and not any less efficient way you can learn from the FTN95 example, see multithreading folder on your harddrive. That's really very cool, specifically combined with Clearwin, and unlike MTASK, is native in FTN95 where developers also solved the conflict problem with simultaneous I/O. In some tasks with small synchronization overhead you get almost proportional to number of cores speedup. Now we have 8-12 core Intel/AMD processors, soon will have ones with 48-cores etc... Parallelization is the most probable road ahead for many years.

I hope may be someone will make libraries and examples for computing on GPUs. As you can see above in some cases speedups on "gaming computers" are very high and iall that's orders of magnitude cheaper then on CPUs