Re: disappointing CUDA speed
- To: mathgroup at smc.vnet.net
- Subject: [mg114217] Re: disappointing CUDA speed
- From: Fred Klingener <gigabitbucket at BrockEng.com>
- Date: Sat, 27 Nov 2010 03:38:00 -0500 (EST)
- References: <iclfde$l9u$1@smc.vnet.net> <ico240$nhj$1@smc.vnet.net>
On Nov 26, 5:28 am, David Koslicki <dmkosli... at gmail.com> wrote:
> On Nov 25, 5:56 am, Gianluca Gorni <gianluca.go... at fastwebnet.it>
> wrote:
>
>
>
> > Hi,
>
> > I have a 1 year old Apple MacBookPro. I installed
> > the cudadriver_3.1.17_macos and then tried the first
> > examples in the documentation:
>
> > Needs["CUDALink`"]
> > CUDAQ[]
> > True
> > randM = RandomReal[1, {3000, 3000}];
> > AbsoluteTiming[randM.randM;]
> > {2.688389,Null}
>
> > AbsoluteTiming[CUDADot[randM, randM];]
> > {7.328353,Null}
>
> > Quite a letdown.
> > Did I do something wrong?
The casting around I've done persuades me that the native Dot[] must
be the most streamlined, compiled, parallelized, optimized Mathematica
function already, and it'll probably take a huge calculation on the
GPU to pay for the overhead.
I think there will probably be more potential for gain in the more
complex parts of the linear algebra library.
Seems to me it would be a good business proposition for Wolfram to
include OpenCL and CUDA linear algebra libraries and pick the best one
for the installed hardware. There's a lot of FUD surrounding this
(just search this newsgroup for "CUDA"), and a few targeted benchmarks
would go a long way toward building user confidence, selling a lot of
new hardware and solving a lot of sophisticated problems in math,
physics and engineering.
Maybe these could fill a section in my other favorite unrealized
Wolfram business proposition - $0.99 iPaclets. (The iPaclet store
would poll your machine hardware, drivers and software and make
recommendations.)
Cabin fever dreams?
Fred Klingener