Memory bandwidth is the rate at which data moves between a processor and memory. In GPU kernels, it can become the bottleneck: how threads access global memory may matter more than how much computation they perform. section 10.0
When more threads do not help
A kernel launched with 10,000 threads is correct and well organized, yet barely faster than its CPU counterpart. The missing speedup points to how its threads read memory. section 10.0



