How to define Hardware requirements for CUDA apps ? How to define hardware requirements to run my Cu

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

We are starting the development of a CUDA application and we need to define the minimum hardware specs for this project.
Today, our clients uses Windows XP and Vista, on Thinkpad T61p (Quadro FX570M 256m), Dell M65( Quadro Fx350 256mb), and some desktops equipped with NVidia series 8. But other clients may buy/use low end graphics, as nv 8400m with 128mb.
Our datasets range from a few megabytes up to ~150Mb of raw data, and will be used to perform CUDA BLAS operations, as well in house developed math procedures.

So ask :
a ) Is possible to know how much video card memory is required before processing a dataset (CUBLAS) ?
b ) What is the minimal video card memory for CUDA / CUBLAS for single video card system, as notebooks ? In another words, is 128mb enough ?
c ) Is Turbo Cache memory used by CUDA ?

Thanks
Alessandro

2

I don’t think the CUBLAS functions do in-place calculations, so you’re looking at whatever matrix size you use times 2 (ex: single-precision 100x100 matrix uses 4100100=40000 bytes, then multiply that by 2). Video card memory has nothing to do with actually having CUDA capabilities in the card…check the nVidia website for compatible cards; anything listed on there is fine, as long as it has enough memory to hold your datasets.

3

No. Most of the CUDA SDK samples don’t even run on 128MiB cards. CUDA reserves ~50 MiB for unknown uses, plus you have the graphics buffers and what not, leaving very little left for even simple CUDA applications.