cuda - Does multiProcessorCount gives the number of streaming multiprocessors? -


i hardly understand value given multiprocessorcount property represent, due fact experience difficulties in grasping cuda architecture.

i'm sorry if of following statements appear naive. understood far, here hardware "layers":

  • a cuda processor grid of building blocks.
  • a building block composed of 2 or more streaming multiprocessors.
  • a streaming multiprocessor composed of many streaming processors, called core.
  • a streaming processor "massively" threaded, meaning implements many hardware managed threads. 1 streaming processor, 1 core, can compute 1 thread @ time, has many "hardware threads" can load data while waiting turn computed sp.

on software side:

  • a block composed of threads, , executed streaming multiprocessor
  • if 1 launched more blocks number of streaming multiprocessors on card, guess blocks wait in sort of queue, executed.
  • software threads distributed streaming processors, distribute them hardware threads. , similar previous case, if 1 launched more threads streaming processors can handle hardware threads, software threads wait in queue.

in both cases, max number of threads, , blocks, allowed launch, independent number of streaming multiprocessors, streaming processors, , hardware threads of each streaming processor, exist on card. notions software!

am @ least close reality?

with being said, multiprocessorcount property gives? on 610m, says have 1 multiprocessor... mean have 1 streaming multiprocessor? have building block composed of 1 streaming multiprocessor? seems impossible me. , mean can execute 1 block @ time! besides, when specifications of card says have 48 cuda cores, talking streaming processors?

perhaps this answer help. it's little out of date since refers old architectures, principles same.

it entirely possible gpu consist of single sm (streaming multiprocessor), if mobile gpu. single sm, composed of multiple cuda cores, can accommodate multiple thread blocks (up 16 on latest kepler-generation gpus).

in case, 610m gpu has 1 streaming multiprocessor (sm), composed of 48 cuda cores (aka streaming processors, sps).


Comments

Popular posts from this blog

SPSS keyboard combination alters encoding -

Add new record to the table by click on the button in Microsoft Access -

javascript - jQuery .height() return 0 when visible but non-0 when hidden -