Texture memory vs. constant memory access latency

[fred_em]

Hi,

I am unclear about the latency of Texture memory vs Constant memory accesses with CUDA. I am only interested in accesses when the cache is hit.

Section §5.3.2.5 of the CUDA Programming Guide, about Texture Memory, says:

Reading device memory through texture fetching present some benefits that can

make it an advantageous alternative to reading device memory from global [u][b]or

constant memory[/b][/u]:

If the memory reads do not follow the access patterns that global or constant

memory reads must respect to get good performance (see Sections 5.3.2.1 and

5.3.2.4), higher bandwidth can be achieved providing that there is locality in the

texture fetches (this is less likely for devices of compute capability 2.0 given that

global memory reads are cached on these devices);

The following document about the GTX280 http://www.networkmultimedia.org/Publications/practicals/beyer2009.pdf (chart at the top of PDF page 25) mentions a very small latency (~register latency) for Constant memory accesses when I hit the cache, and a latency of 100 cycles when hitting the Texture cache.

Is it yes or no much faster to access the Constant memory cache (~register latency) than it is to access the Texture Memory cache (sounds like this is 100 cycles)? Is it the same for G80 and Fermi boards?

Any help would be appreciated.

Cheers,

Fred

[tera]

Yes. Check out this paper: Demystifying GPU Microarchitecture through Microbenchmarking. You’ll have to redo the measurements for Fermi though.

[schmichael]

Yes. Check out this paper: Demystifying GPU Microarchitecture through Microbenchmarking. You’ll have to redo the measurements for Fermi though.

Cool, thanks.

It just happens that I am working with a C1060 :)

[schmichael]

Yes. Check out this paper: Demystifying GPU Microarchitecture through Microbenchmarking. You’ll have to redo the measurements for Fermi though.

Cool, thanks.

It just happens that I am working with a C1060 :)