This project is read-only.

CUDAfy.NET and GTX680 ?!

Mar 25, 2012 at 6:21 PM

 

Hi,

Someones have make the jump to the latest Nvidia GTX680 using CUDAfy ?   If we see a big difference?   Because looking at the specs, the 1538 CUDA cores at 1Gz looks very interesting ... but depending the use :

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/1

http://www.anandtech.com/Show/Index/5699?cPage=25&all=False&sort=0&page=18&slug=nvidia-geforce-gtx-680-review

 

Regards,

Mar 26, 2012 at 10:50 AM

Yeah it does look interesting at first glance, however compute performance looks like it could be a step backwards for many algorithms.  Maybe a new CUDA compiler will be necessary to optimize for the new architecture.  I would not be in a rush to get one.

Mar 26, 2012 at 4:23 PM

 

I thanked the abstraction was done at driver layer ... Allowing to keep our current CUDA codes "as is" with the Kepler arch.  May be I'm wrong, but I will try to verify again that. 

Dec 3, 2012 at 12:38 AM

Is there any more info on this 8 months later?

How do you find out about compute performance for each CUDA card?

Are there any comparisons of nvidia cards for CUDA performance? 

If I only want single-precision performance which card is best?

Dec 3, 2012 at 1:08 AM

I had a 570 gtx and now a 680 gtx. I see aprox 20% improvement for my single-precision code (pretty heavily optimised, too). Can't say I'm too happy. I guess my numbers are probably very dependent on the particular code I run. Yours may differ.

But you do get more registers to fool around. One thing the new ones are good at tho, is doing reductions within a warp (__shfl_xor).

 

Dec 3, 2012 at 2:50 AM

If I compare the specs of those 2 processors on Wikipedia:

http://en.wikipedia.org/wiki/GeForce_500_Series

http://en.wikipedia.org/wiki/GeForce_600_Series

 

The 570 is 1405 gflops, and the 680 is 3090 gflops - so clearly gflops is not the comparison I should be looking at. Is there some other spec that will help?

 

Dec 3, 2012 at 9:01 AM

Compared to Fermi, I believe the status for Kepler is still better single point precision performance and poorer double. The only Kepler device where this is not the case is the top of the range K20.