I am looking to setup a system for developing prototype applications with CUDA. Can I just buy any of the low end NVidia graphics cards, and replace the graphics card on my desktop? Looking at [url=“CUDA GPUs - Compute Capability | NVIDIA Developer”]http://www.nvidia.com/object/cuda_learn_products.html[/url], there is quite a wide range of choices, and I’m confused. Could I request your suggestions for setting up my first CUDA system.
If you want decent performance, double precision support and CUDA Capabilities 1.3, then go for a GTX 260.
If you want decent performance and single precision is enough, go for a nVidia 8800GT or 9600 GSO (the 384MB or 768MB models).
If you need silence during development and prototyping, go for a passively cooled 8400GT or 8500GT. You can always run the production version on high end hardware later.
Don’t pick the dated G80 achitecture (8800GTS), prefer G92 models over G80 due to the added global atomics support which is useful in many situations.
Make sure you`ve got enough RAM on the graphics chip. 768 MB or 1GB is fine.
Pick a decent motherboard with PCI Express slots, possibly several 16x slots with double spacing between them, so you could run two or more GPUs in parallel.
Well, personally I’ve bought 8400GS (super low end) for CUDA development, and it’s fine, I’m happy with passive cooler :-)
9600GSO would offer good performance/$
How many processor cores are on the 8400 GS? I couldn’t find this info on the NVidia website. I’m primarily buying this for developing parallel CUDA applications, rather than for Graphics support.
Regarding cooling, how do I tell from the NVidia product page whether a card is passively cooled? Does the 9600 GSO get very noisy? Do these products need to be used in a room with air conditioning?
there are both versions of the 9600 gso available, actively and passively cooled. just look, what you can get in a store/at an e-tailer.
the nvidia product pages only show the reference designs.
a 8400gs has 16 shaders at 900 mhz (=nothing) and a very slow ddr2 memory with a 64bit interface.
also make sure, that you don’t need any features introduced with higher compute capabilities (see programmer’s guide). especially the changed coalescing rules, atomic operations and double precision may be points to consider.
personally, i would go for a gtx260 at the moment.