Weird behaviour when modifying architecture in build rule
Hello!

I got a mysterious behaviour while modifying the architecture in Build rule in VS2008. I am using a GTX285 (architecture 1.3) and execution is twice longer using sm_13 than using sm_11. It should be better when selecting the right architecture.

The kernels perform mainly texture hits and global store with few computations.

Does somebody get an idea on where it comes from?

Thanks,
Sam
Hello!



I got a mysterious behaviour while modifying the architecture in Build rule in VS2008. I am using a GTX285 (architecture 1.3) and execution is twice longer using sm_13 than using sm_11. It should be better when selecting the right architecture.



The kernels perform mainly texture hits and global store with few computations.



Does somebody get an idea on where it comes from?



Thanks,

Sam

#1
Posted 10/26/2010 04:31 PM   
Hello!

I got a mysterious behaviour while modifying the architecture in Build rule in VS2008. I am using a GTX285 (architecture 1.3) and execution is twice longer using sm_13 than using sm_11. It should be better when selecting the right architecture.

The kernels perform mainly texture hits and global store with few computations.

Does somebody get an idea on where it comes from?

Thanks,
Sam
Hello!



I got a mysterious behaviour while modifying the architecture in Build rule in VS2008. I am using a GTX285 (architecture 1.3) and execution is twice longer using sm_13 than using sm_11. It should be better when selecting the right architecture.



The kernels perform mainly texture hits and global store with few computations.



Does somebody get an idea on where it comes from?



Thanks,

Sam

#2
Posted 10/26/2010 04:31 PM   
Does your kernel use double precision? If yes, it would be downgraded to single precision with sm_11 and performed for real with sm_13
Does your kernel use double precision? If yes, it would be downgraded to single precision with sm_11 and performed for real with sm_13

#3
Posted 10/29/2010 02:11 AM   
Does your kernel use double precision? If yes, it would be downgraded to single precision with sm_11 and performed for real with sm_13
Does your kernel use double precision? If yes, it would be downgraded to single precision with sm_11 and performed for real with sm_13

#4
Posted 10/29/2010 02:11 AM   
Scroll To Top