NEWBIE:max size of shared memory of a block?

Accelerated Computing CUDA CUDA Programming and Performance
1

sorry i konw it’s a newbie question.i’ve google it for a while and got no answer.
16384B shared memory for a MP in 1.3, but how about a block?

2

It’s the same value. A multiprocessor can run between 1 and 8 blocks simultaneously, depending on how many resources each block uses. If you use 16000 bytes of shared memory per block, then only one block will run at a time on each multiprocessor.

Note that in practice you can’t use all 16384 bytes of shared memory, because the compiler reserves some of it (256 bytes?) for passing kernel parameters and grid/block information.

3

There’s one thread on these forums confirming that using an unsupported hack you CAN use all 16384 bytes (definitely not tmurray approved)

Christian

4

thanks.