Stefan Evert wrote: > > On 19 Nov 2008, at 07:56, Prof Brian Ripley wrote: > >>> I just read an announcement saying that Mathematica is launching a >>> version working with Nvidia GPUs. It is claimed that it'd make it >>> ~10-100x faster! >>> http://www.physorg.com/news146247669.html >> >> Well, lots of things are 'claimed' in marketing (and Wolfram is not >> shy to claim). I think that you need lots of GPUs, as well as the >> right problem. > > Which makes me wonder whether R would be able to make use of this > processing power, since the figures claimed by Wolfram are very probably > for single-precision floats? (Am I right in thinking that R only works > with double precision?) > > According to the nVidia Web site, the Tesla architecture is _ten times_ > slower for double-precision operations than for single-precision, which > makes it seem far less amazing than at first sight. >
It wouldn't be rocket science to add a single-precision data type to R, it is just a large amount of work to get all details right, and noone has found it worth the trouble. Probably more of a stumbling block is that GPUs traditionally have had a rather cavalier attitude to floating-point standards. We have had our share of problems with compilers that breaks the rules (or users setting flags like -fast-math that allow rule breakage), and ensuing breakage of user code and/or "make check", so I don't think it is viable to try and get the GPU to take on all of R's computations. Special-purpose computation, e.g. an R package interfacing to "libgpublas.so" or whatever, is always a possibility, but that really goes without saying. -- O__ ---- Peter Dalgaard Øster Farimagsgade 5, Entr.B c/ /'_ --- Dept. of Biostatistics PO Box 2099, 1014 Cph. K (*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918 ~~~~~~~~~~ - ([EMAIL PROTECTED]) FAX: (+45) 35327907 ______________________________________________ [email protected] mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
