more -------- Original Message -------- Subject: Re: [Jchat] Parallel Processing for J Date: Fri, 28 Feb 2014 10:58:39 -0500 From: Robert Bernecky <[email protected]> To: [email protected] CC: Sven-Bodo Scholz <[email protected]>
SAC, which I mentioned yesterday, includes back end support for CUDA, which makes the task of programming for CUDA MUCH easier than playing ironman-I-can-code-anything: just specify -target cuda when you compile the code The SAC web site, sac-home.org should include several papers on the subject, but I only see one now, on Modelica. Speedups that I have seen are relatively impressive, given a reasonable piece of tin under the CUDA card. Perhaps Sven-Bodo Scholz can provide some performance numbers and/or other papers on the subject. WRT the K1, bear in mind that lots of parallel cores do not a supercomputer make, because you also need gonzo memory bandwidth and, typically, gonzo main memory for realistic problems, neither of which exist in contemporary cell phones. Do not be misled by the "same architecture" marketing hype. It may be true, but a dinky ferry boat and an ocean liner are also built using the "same architecture": we do not expect the same performance characteristics from them. It would certainly be interesting to see how well some serious physics models would work on a CUDA-enabled cell phone. (I have visions of melting plastic and two-minute battery life, despite Nvidia's claims of "low power". ) Bob On 14-02-28 03:34 AM, Skip Cave wrote:
I participated in an ACM webinar today presented by nVidia, where they covered their latest multiprocessor "accelerator" GPU chip, the Tegra K1<http://bit.ly/1mJYL5O>. The K1 has 192 CUDA parallel cores as well as four ARM-15 cores. It is designed to be used in cellphones, so it uses very low power. The webinar link is at: http://bit.ly/1cq6E6E I would recommend taking a look at the presentation, as it shows just how far we have come in putting serious multiprocessing power in the hands of the masses nVidia claims the architecture is being used in various "green" supercomputers, running models of Molecular Dynamics, Quantum Chemistry, Material Science, Weather& Climate, Lattice QCD, Plasma Physics, Structural Mechanics, and Fluid Dynamics. They claim that they have built compilers and interpreters for their multiprocessing architecture in Fortran, C, C++, Pyton, and F#, which will make it easy for programmers to distribute processing load across the array of CUDA cores in the K1, as well as the four ARM processors. I always felt that the primitives in J define a powerful and general set of array operations that could benefit from a parallel processing architecture. The nVidia processors will be showing up in cellphones in the near future. It would be interesting to see if J could take advantage of the amazing parallel processing power of these chips - a supercomputer in your pocket. Skip. http://bit.ly/1cq6E6E Skip Cave Cave Consulting LLC ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
-- Robert Bernecky Snake Island Research Inc 18 Fifth Street Ward's Island Toronto, Ontario M5J 2B9 [email protected] tel: +1 416 203 0854 ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
