Hard to tell without serious benchmarks to support any claims.
Oh, and a hugh amount of money to acquire that extensive variety of cards!
assuming they are equally supported MPE cards
core clock hands down with some comments. obviously a 64 meg memory card would not be very good either vs 1024Meg
where a 512meg could beat a 1024 if its clock is faster.
a 480 will best a 285 and any present quadro. of course the 480 is not presently a supported card so limited.
a 285 beats all Quadros save the 5800 (which is a 285)
What about 480 vs Sparkle GT 220 with 2GB memory?
You compared 480 and 285, but 480 have higher/faster every parameter (memory and core clocks, amount of memory, memory standard, etc.)...
a 220 is a joke no matter how much ram, minimum IMHO is a 260 (what most quadro's are made from)
Hmmm.... What about GTX 480 with 1,5GB memory vs GTX 295 with 1,7GB memory?
reread my post
1 person found this helpful
Due to how Cuda works and memory management is completely up to the application designer when they write to Cuda it will differ somewhat. But the main specs are the number of Cuda Cores and the Ram bandwidth. Really the Ram bandwidth is higher on the list due to the CPU having to constantly burst data down to the video card ram and the ram GPU returning the favor after GPU processing. The faster the Ram bandwidth the faster the GPU clears the threads and gets on with the next threads. In regards to the GPU, consider the video ram as the same as CPU cache when dealing with Cuda. So the GPU is constantly sending data to the GPU ram and pulling from it. Also the GPU is sending data from the GPU Ram to the system Ram. This puts the GPU ram as the transportation hub so to speak. How much faster will you get to your destination if you are traveling 100MPH instead of 50MPH. The concept here is not much different other than you also have an 8 lane highway instead of a 4 lane highway so there is also not as likely a chance for a traffic jam. That is why the interface width is important as well. So the order is:
Memory Bandwidth/Cuda Cores(could be reversed depending on memory management in application)
Memory Frame buffer size(ie amount of ram physically on the card)
Great thanks for everyone for help.
One more question. For example in host I have few CUDA cards. I'd like to choose with one should by used by Adobe Premiere for GPU computing. Is that possible?
With the Beta driver they just put up not long ago, yes. Before then it was harder.
Eric, I assume that you are talking about an nVidia beta driver, is that correct?
I tried my three monitor experiment where I had Premiere spread across two monitors from the GTX 285 and on an old FX 1700 I had a 1920 x 1080 monitor. I setup playback on the 1920 x 1080 monitor and it came out as an approximately 1200 x 1080 image (horizontally compressed) I tried all three aspect ratio adjustments with no changes.. I tried an old nVidia 6200TC card and got the same results. Do you think that driver may help?
Yes it adds an option in the Nvidia control panel to select the Cuda GPU.
I have installed it here and the option is there. I however have not tested it will more than 1 card. You will likely have an issue trying a 200 series card and a older gen card but doesn't hurt to try.