Am just getting started on building a development system for Cuda…can i use one of the many nvidia graphic cards that are manufactured by vendors such as xfx, asus, pallit etc etc. ? There are many out there and some have overclocking / different memory sizes etc
Does anyone have any specific recommendations on which cards work well / are cost effective etc
Regards Peter G
Are you aiming for a high end card or lower end card? If you are going to be working on problems that require lots of device memory, the 8800GT 1GB version is pretty solid and a fairly good price. That’s what I got since I am working with matrices that are large (15000x15000) which is pushing 860mb’s if I want to store it on device all at once (which I do). You can use any card that is nvidia 8 series or above. And it can be any brand, but I recommend a good name brand like evga or one you are comfortable with.
my application area is signal processing …i need high computional power/throughput but the amount of static data needing to be held will be modest ( 8kx8k max ) The graphics i need at simple almost static control panes. I need to stream the data in and out via either a high speed lan or fire wire
So i assume that the reputable card supliers have no compatibility issues running cuda based applications ( or should i say …Cuda has no issues using non Nvidia made compatible cards) ?
Peter G
Well, NVIDIA actually doesn’t manufacture any of their own cards so you must buy one from xfx,evga,coolermaster,etc… All of them are supported just fine by CUDA. Just don’t buy an ATI card or you’ll have issues :wink:
I will only add that, in my experience, CUDA applications are very sensitive to overclocking (they crash or the memory gets corrupted) so I would recommend against purchasing a card that comes overclocked from the factory. Find one with the NVIDIA recommended clock speeds.
As far as which card to buy, I find it hard to justify anything but the fastest (currently the GTX 280). You do want your app to run fast right? But them my view is skewed because I’m doing academic research and the cost of a GTX 280 is just a drop in the bucket (that and NVIDIA is very good about giving researchers more hardware then they know what to do with :)
If you want the best performance/dollar, things get complicated. Memory throughput bound code will run best on the 8800 GTX (if you can find one anymore…): you can get 70 GiB/s global memory bandwidth. If you are compute bound, then the 9800 GTX+ or the 8800 GTS 512MB could be your best bet as they can crunch more FLOPS than the 8800 GTX but only offer 50 GiB/s.
But… you mentioned streaming data from an external device so you may be PCIe bound and thus you wouldn’t want a 8800 GTX since it is only PCIe 1.0. If you need to maximize the transfers over PCIe, the best performance can be had by any NVIDIA card newer than the 8800 GTX and an Intel chipset MB.
As far as which card to buy, I find it hard to justify anything but the fastest (currently the GTX 280). You do want your app to run fast right? But them my view is skewed because I’m doing academic research and the cost of a GTX 280 is just a drop in the bucket (that and NVIDIA is very good about giving researchers more hardware then they know what to do with :)
If you want the best performance/dollar, things get complicated. Memory throughput bound code will run best on the 8800 GTX (if you can find one anymore…): you can get 70 GiB/s global memory bandwidth. If you are compute bound, then the 9800 GTX+ or the 8800 GTS 512MB could be your best bet as they can crunch more FLOPS than the 8800 GTX but only offer 50 GiB/s.
But… you mentioned streaming data from an external device so you may be PCIe bound and thus you wouldn’t want a 8800 GTX since it is only PCIe 1.0. If you need to maximize the transfers over PCIe, the best performance can be had by any NVIDIA card newer than the 8800 GTX and an Intel chipset MB.
[snapback]412636[/snapback]
[/quote]
Thanks for the inputs so far…am tempted to get the gtx at this point… what’s the situation in regard the various instruction sets…i would like to have the double precision math support…v1.2?..is this available in the 9x series and the gtx? or is it simply a waiting game for the newer cards to get that support ?
Peter G
Double precision is supported in the GTX 280 and 260, but none of the earlier cards. (The caveat is that double precision is around 8 times slower than single precision, so you definitely want to scale your expectations by that. You can mix single and double precision, of course, for better performance.)
I’m quite partial to the 512MB 8800 GTS right now, for my problem I see a significant speed boost at little extra cost, but I think other cards are becoming a smarter buy.
Keep an eye out for GTX 260 prices.
I’ve used eVGA, XFX and PNY, 8800 GT and GTS. All were fine.
Thanks for the inputs so far…am tempted to get the gtx at this point… what’s the situation in regard the various instruction sets…i would like to have the double precision math support…v1.2?..is this available in the 9x series and the gtx? or is it simply a waiting game for the newer cards to get that support ?
Peter G
[snapback]412644[/snapback]
[/quote]
If you want double, then get the 260 or 280, since you don’t have much other choice :-P.
I have to say that nearly any CUDA developer wants a GTX board… 260 is fine. The GTX architecture is so much superior for CUDA… it’s clear that it was optimized for GPGPU.
You get more registers, much better coalescing, more threads in flight, double and long long support, better bandwidth, PCIe 2.0, shared atomics… it’s just a big difference.
The 260 is fine, the 280 is faster but the architecture is the same for both.
And it looks like the recent price cuts have percolated to the online vendors.
The GTX 280 can now be had for $450:
http://www.newegg.com/Product/ProductList…rds&Order=PRICE
The GTX 260 is now $300:
http://www.newegg.com/Product/ProductList…rds&Order=PRICE
The video game market is brutal.