CUDA: Difference between revisions

Line 79: Line 79:
* Supports some maximum number of blocks and threads (~8 and ~768 on G80).
* Supports some maximum number of blocks and threads (~8 and ~768 on G80).
====Block sizing====
====Block sizing====
'''FIXME: review all this!'''
How tightly can we bound the optimal block size '''T''', given a warp size ''w''?
How tightly can we bound the optimal block size '''T''', given a warp size ''w''?


Line 98: Line 99:
** relative speedup tends to 0 as work grows arbitrarily on finite SMs
** relative speedup tends to 0 as work grows arbitrarily on finite SMs
** relative speedup tends to ''1/Frac<sub>par</sub>'' on infinitely many SMs
** relative speedup tends to ''1/Frac<sub>par</sub>'' on infinitely many SMs
'''FIXME: review all this'''


===Stream Processor===
===Stream Processor===