Table 4.

Profiling results of the intratask implementations with synthetic data set 26 (warp* stands for the improved warp based).

	Performance limitation	Shared memory bandwidth, GB/s	Registers per thread	Shared memory per block, bytes	Theoretical occupancy, %	Achieved occupancy, %
Naive	Memory bandwidth	2213	32	2124	100	99.8
Tile = 2	Memory bandwidth	2005	39	3660	75	74.9
Warp	Instruction and memory latency	362	32	6540	10.9	10.9
Warp*	Compute	262	44	2000	62.5	62.5