It's just a buzzword. A trillion floating-point operations per second? What operation? For example, floating-point division is much slower with much worse throughput than addition subtraction and multiplication.

Architecture also matters a lot. Remember when Sony said the PS3 GPU had 1.8 TFLOPs? Really? So surely the PS3 GPU was roughly as fast as the PS4 GPU which had 1.84 TFLOPs right? It's bullshit. FLOPage comparisons are really only meaningful if the two processing units being compared have the same architecture. So even if the PS5 has 9 TFLOPs or 12 or whatever, that number is not directly comparable to the FLOPs of the PS4 and PS4 Pro for example. Not even the FLOP numbers of AMD and Nvidia cards can be meaningfully compared. So saying that the 12 TFLOP XSX GPU is equivalent to a Nvidia 2080 Super just based on that number is also bullshit. The only way to see how it compares is to benchmark it.

Agree. Flops is mostly meaningless since it is so disconnected from what the workload is.

Traditionally, flops used to be a measure of how many floating point multiplications you could do.

As a compromise between the much cheaper additions and the much more expensive divisions (not to mention logarithms, ...)

At least in the hpc and super computer world it is completely meaningless and will just make people shake their heads if you start talking about it.

Now that flops are important selling points to low end end user consumer hardware things got even worse,

it would not surprise me at all if shitty vendors have re-defined it to be how many additions they can do instead.

Just so they can inflate a meaningless number and impress ignorant consumers and fanbois.

For real serious compute power on big-compute, the real measure that people actually care about is much more domain specific.

For example, for math, matrix multiplications, vectors, fem computations, etc

no one cares about a meaningless flops number. Here you would have a different domain specific measure called "daxpy".

Other types of compute intensive workloads have other types of domain specific measures of compute power.

In all of these domains, the flops number is completely meaningless.

The only people that care about flops are laymen and fanbois that do not understand it is a meaningless metric.

It is a meaningless marketing term to fool people that do not know better.