Complete Newbie Checklist Can someone confirm/correct these Qs...

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I would like to learn a little about GPGPU, with the possible aim of rewriting some software we use at work (basically lots of cross-correlation). The idea is that I do some first experiments on my own and then, if it looks promising, work would probably pay for a decent setup. So at this point I want something cheap and cheerful.

Is the following correct (these are pretty basic, I know):

1 - Only a single graphics card is necessary (ie I can use the card driving the display while it is “idle”)

2 - The ASUS card EN8400GS/HTP/256M will work OK with CUDA (described at [url=“http://www.bip.cl/ecommerce/buscador2.php3?id_producto=2470”]http://www.bip.cl/ecommerce/buscador2.php3?id_producto=2470[/url] - I live in Chile, so not all hardware is available)

3 - I can use OpenSUSE 11, following the instructions here [url=“SUSE – Open Source Lösungen für Enterprise Server und Cloud | SUSE”]SUSE – Open Source Lösungen für Enterprise Server und Cloud | SUSE (note that OpenSUSE 11 doesn’t seem to be in the “supported” list for CUDA).

Thanks very much. Also, is there an IRC chat or similar for asking very basic questions like this? Or some newbie-friendly forum? I looked in various FAQs and (1) which seems like the most basic FAQ of all, didn’t seem to be answered anywhere. :o(

Cheers, Andrew

2

1 - Yes, a single card is fine. The NVIDIA driver will multiplex graphics operations by X11 and CUDA kernel calls. However, the X11 display cannot update while a kernel call is running (not the entire program, just each call to a global function on the GPU). If your kernel calls take less than 5 seconds, this is just an interface annoyance. Kernel calls longer than 5 seconds will be prematurely terminated (called the “watchdog timeout”). You can avoid the watchdog by not running X. Then there are no time limits.

2 - That card should certainly run CUDA. 256 MB is definitely the lowest you would want to go for CUDA development, as some of the memory is taken for the display. How much data do you think you’ll need to load on the GPU?

One warning about the 8400, though. The memory bandwidth and shader count (16) is so low, you might find that your CPU easily beats the GPU in your testing. The 8400 is sufficient to demonstrate CUDA working on real hardware, though.

3 - I don’t use SUSE, so hopefully someone else can answer this.

3

Thanks! Showing it works is enough for now (I hope!).

I guess openSuse 11 support will arrive soon, or I can use rpms 10.3.

Tomorrow I’ll pick up a cheap card…

Cheers,
Andrew