[ltp] cuda and the quadro nvs 140M, memory allocation issues

[Micha Feigin]

On Sat, 10 Jan 2009 19:27:48 +0000
Richard Neill <rn214@hermes.cam.ac.uk> wrote:

> 
> 
> Micha Feigin wrote:
> > I'm trying to learn cuda programming (run programs on the gpu) under my t61
> > with the quadro nvs 140m card.
> > 
> > I keep getting memory allocation failures and anything but very small
> > allocations.
> > 
> > I don't know if it's related but it seems that the card is wrongly reporting
> > the amount of memory it has.
> > 
> > AFAIK it should have 256MB memory. nvidia settings shows though that it has
> > 512MB and I don't know how to check what it's saying.
> > 
> > I'm also guessing that X is taking too much of the video memory but not
> > sure.
> > 
> 
> At a guess, your card is pulling the common trick of claiming a large 
> chunk of regular system RAM, and using it as a (lower-bandwidth) 
> addition to the card's main RAM; you can usually switch this off in the 
> BIOS.
> 

That's what I was guessing but there is not bios option to play with this as
far as I could find. I also couldn't see a module option to change this.

looking at free and /proc/meminfo it seems that there are 53.75 MB missing,
which is 1/5 of what is supposed to be missing but is still weird. (machine
should have 2GB installed, which is 2097152 bytes. meminfo says there are only
2042112 bytes)

> You probably need
>      1600x1200  pixels
> x   2 screens (if dual display)
> x   4-bytes per pixel for 32-bit RGBA colour
> x   2 for double-buffering, if you want 3D accel
> 
> =  32Mbyte minimum for your actual system. So I'd expect you can 
> allocate 128M safely for your experiments.
> 
> 
> Richard
> 
> 
> P.S. For amusement...don't actually do this:
> http://thedailywtf.com/Comments/The-Skys-the-Limit!.aspx?pg=2
> 
> 
> 
> 
>