Code snippet measures integer operations, text mentions floating point operations.

algorithmica/content/english/hpc/cpu-cache/bandwidth.md

Line 30 in 990b6b1

    
           - When the whole array fits into the lowest layer of cache, the program is bottlenecked by the CPU rather than the L1 cache bandwidth. As the array becomes larger, the overhead associated with the first iterations of the loop becomes smaller, and the performance gets closer to its theoretical maximum of 16 GFLOPS.

Not sure if the meaning of FLOPS has been extended in practice, or integer op throughput is tightly coupled to flops.

It's an issue that needs to be fixed worldwide, not just in this book. People use GFLOPS to refer to anything. There is a more general GOPS term, but it sounds ridiculous.