4 posts.
A warp's 32 loads are served in 32-byte sectors. Neighboring threads reading neighboring words need 4 sectors; a large stride needs 32, wasting 87.5%.
In SIMT, the 32 threads of a warp issue one instruction together. When they branch different ways, the warp runs each path in turn with lanes masked off.
When the work is parallel, does enough arithmetic per byte, and is big enough to pay for the data transfer. A roofline comparison, worked through.
A GPU splits work into thread blocks, runs each on a streaming multiprocessor in warps of 32, and hides memory latency by switching warps.