Reply To: T-shirt designers unite and take over

[texaslabrat]

Breakable wrote: Eric mentioned that his simulation cannot be paralelised.
I am guessing INCITE can be only implemented by a lot of parralel computer cores.
So there is no speed up by running a serial task there, because you could use just one computer core.
Its better to run on a GPU which has more serial powe.
Of course dont forget that most GPU’s are also implemented by using several cores in paralel,
so you should choose one that has the fastest core, instead of the most teraflops.

While I don’t want to speak for Dr. Lerner, I think you’ve mis-interpreted his statement. He said nothing about the problem not being a highly parallel problem….rather he mentioned that it is the type of problem that doesn’t easily lend itself to being farmed out to large numbers of unlinked processors such as in the folding@home or distributed.net projects (which have the luxury of divvying up work to be completed asychronously since the results from one data set do not depend on another for the most part). That, in itself, means nothing in regards to the parallel nature of the problem, but rather that the steps involved in the calculation are closely related to one another and can’t be done asynchronously. And that makes perfect sense…I would imagine that the calculations are trying to keep track of a bunch of particles and force lines…and every step changes the environment for all of them. Thus, this is the quintessential parallel problem insofar that all the calculations must truly be parallel in a strict sense (all the forces, particle vectors, etc must be calculated and results placed into a multi-dimensionsal array before moving onto the next time slice, which will use that array as the initial conditions for the next turn of the crank). It would seem that the highly parallel nature of gpu architecture would be very well suited, and TFLOPS very much matters 😉 That said, the ease of programming is probably the biggest variable…I’ve heard that CUDA is much more “programmer-friendly” than ATi’s CTM api….but I have no first-hand experience in programming either of them.

I look forward to learning more details whenever Dr. Lerner has an opportunity to write something up.