Cheapest GPU supporting latest CUDA features?
April 30, 2011 12:18 PM Subscribe
What is the minimal Nvidia GPU that supports all the newest features for CUDA programming?
I am building a new PC, and I want to put in a Nvidia GPU so I can learn some CUDA programming. I have heard that the latest generation has some major advances like unified address space and full C++ support. That sounds pretty good to me! However, a lot of the information I can find on GPUs targets gamers. I don't care about what settings I can use on Crysis; I just want to write cool programs. What is the cheapest Nvidia GPU I can get that supports all the good programming features?
I am building a new PC, and I want to put in a Nvidia GPU so I can learn some CUDA programming. I have heard that the latest generation has some major advances like unified address space and full C++ support. That sounds pretty good to me! However, a lot of the information I can find on GPUs targets gamers. I don't care about what settings I can use on Crysis; I just want to write cool programs. What is the cheapest Nvidia GPU I can get that supports all the good programming features?
Best answer: You want a card that has the highest available "compute capability" - currently this is Compute Capability 2.1.
This section of Wikipedia should give a pretty conclusive answer, pending that it's completely up to date: CUDA - Supported GPUs.
Generally, one wants the cheapest GTX models of the highest Compute Capability - currently, the GTX 460 and 560 (and probably the 550, which I haven't had a look at). The 560 Ti is a lot of fun to work with, crazy fast, total powerhouse. It's hard to go wrong with a GTX 460. I'm not as absolutely confident in the GTS 450, but I'm still confident enough that I'd buy it to see.
Then there's the GT 420, 430 and 440. They weren't suited to my purposes when I was last buying these things, but the could very well be what someone in your situation is looking for. As long as the number of CUDA cores is reasonable, e.g. 1/2 rather than 1/10th of the midrange cards, these could be the ticket.
CUDA can be a lot of fun! I hope you'll enjoy it.
posted by krilli at 12:56 PM on April 30, 2011
This section of Wikipedia should give a pretty conclusive answer, pending that it's completely up to date: CUDA - Supported GPUs.
Generally, one wants the cheapest GTX models of the highest Compute Capability - currently, the GTX 460 and 560 (and probably the 550, which I haven't had a look at). The 560 Ti is a lot of fun to work with, crazy fast, total powerhouse. It's hard to go wrong with a GTX 460. I'm not as absolutely confident in the GTS 450, but I'm still confident enough that I'd buy it to see.
Then there's the GT 420, 430 and 440. They weren't suited to my purposes when I was last buying these things, but the could very well be what someone in your situation is looking for. As long as the number of CUDA cores is reasonable, e.g. 1/2 rather than 1/10th of the midrange cards, these could be the ticket.
CUDA can be a lot of fun! I hope you'll enjoy it.
posted by krilli at 12:56 PM on April 30, 2011
(Oh - re. emulation - I believe that emulation is actually being phased out, and isn't supported in the latest more-fun-to-program-against CUDA releases. The reason is chiefly that there are live on-GPU debuggers now available, and emulation mode sucked pretty bad :) )
posted by krilli at 12:58 PM on April 30, 2011
posted by krilli at 12:58 PM on April 30, 2011
Wait, I got lost and forgot to answer the exact question:
What is the cheapest Nvidia GPU I can get that supports all the good programming features?
As per the Wikipedia article, the GT 420 :)
posted by krilli at 12:59 PM on April 30, 2011
What is the cheapest Nvidia GPU I can get that supports all the good programming features?
As per the Wikipedia article, the GT 420 :)
posted by krilli at 12:59 PM on April 30, 2011
Best answer: Arrgh ... Answer-flood. Sorry - I should have looked better at the GT 420. It's an OEM card, so it might be hard to obtain, and it's built with the unusual config of 2GB of VRAM - which says to me that it's a special-purpose card. Doesn't make sense to me for CUDA dev.
So: The GT 430 is probably the realistic cheapest option.
posted by krilli at 1:04 PM on April 30, 2011
So: The GT 430 is probably the realistic cheapest option.
posted by krilli at 1:04 PM on April 30, 2011
« Older I assume this means I can't end up in someone... | New entrant to a Phd programme, any and all tips... Newer »
This thread is closed to new comments.
1. It's possible to emulate CUDA without a GPGPU: "It is possible to develop CUDA software in the absence of a CUDA-enabled GPU. You can test the software in an emulation mode described later in this document." source pdf
2. For $2.10 per hour you can rent a very beefy EC2 GPU virtual machine. Cheaper if you get it as a spot instance.
posted by qxntpqbbbqxl at 12:51 PM on April 30, 2011