Re: CUDA in Octave 322

[Andreas Romeyke]

Hello,

Am Fri, 11 Sep 2009 14:15:04 +0200
schrieb Thomas Ilnseher <address@hidden>:

> Problem here is that you have to  copy data to the GPU's memory nad
> back, call the GPU and so on. This probably gives a reasonable
> overhead, so you'd only accelerate things this way when operating
> with *LARGE* vectors / matrices.

Nope, you would only accelerate things this way when *heavily*
operating on vectors/matrices. The bottleneck is the data transfer
between GPU and CPU. There are limits in branching, there are limits in
jumping or random access the video memory.

Bye andreas

-- 
Software Developer / Dipl. Inform. (FH)
Max Planck Institute for Human Cognitive and Brain Sciences
Department of Psychology
Stephanstr. 1a, 04103 Leipzig, Germany

signature.asc
Description: PGP signature