That’s like saying clock rate and core count are fake terms. Sure, by themselves they might not mean much, but they’re part of a system that directly benefits from them being high.

The issue with teraflops metric is that it is inversely proportional (almost linearly) to the bit-length of the data, meaning that teraflops@8-bit is about 2x(teraflops@16-bit). So giving teraflops without specifying the bit-length it comes from is almost useless. Although you could make the argument that 8-bit is too low for modern games and 64-bit is too high of a performance trade off for accuracy gain, so you can assume the teraflops from a gaming company are based on 16-bit/32-bit performance.

'They are'

It's usually measured as the performance doing a floating point fused multiply-add (fma) - that's it.

But also, multiply then add is the cornerstone of 3D graphics...