Coalesced access to global memory requires threads within the same warp to access contiguous memory addresses; otherwise, a single access is split into multiple transactions, potentially reducing bandwidth utilization to 1/32 of the theoretical peak