Which is faster in CUDA: Constant Memory or Texture Memory?

Question

I know that both is on off-chip DRAM and cached.

But which is faster in access speed? Or in what circumstances one is faster than the other?

Answer 1

Constant memory is optimized for broadcast, i.e. when the threads in a warp all read the same memory location. If they are reading different locations, it will work, but each different location referenced by a warp costs more time. When a read is being broadcast to the threads, constant memory is MUCH faster than texture memory.

Texture memory has high latency, even for cache hits. You can think of it as a bandwidth aggregator - if there's reuse that can be serviced out of the texture cache, the GPU does not have to go out to external memory for those reads. For 2D and 3D textures, the addressing has 2D and 3D locality, so cache line fills pull in 2D and 3D blocks of memory instead of rows.

Finally, the texture pipeline can perform "bonus" calculations: dealing with boundary conditions ("texture addressing") and converting 8- and 16-bit values to unitized float are examples of operations that can be done "for free." (they are part of the reason texture reads have high latency)

Answer 2

Texture memory is optimized for 2D spatial locality (where it gets its name from). You can kind of think of constant memory as taking advantage of temperal locality.

The benefits of texture memory over constant memory can be summarized as follows:

• Spatial locality
• The addressing calculations can be calculated outside of the kernel in the hardware
• Data can be accessed by different variables in a single operation
• 8 bit and 16 bit data can be automatically converted to floating point numbers between 0 and 1.0

See the documentation for more details.