The Cray X-MP 4/16 supercomputer(a.k.a very expensive sofa for the computer lab) I used at uni did ~110megaflops per core.
One class assignment was to write assembler(or fortran) subroutines for matrix multiplications
and keep the pipelines from stalling. How to interleave your data across the memory banks were very important
and to get a passing grade your code had to sustain >95% of theoretical peak performance for these matrix multiplications.
But real men don't measure performance in flops. Measuring it in DAXPY is where the real game is.
Back to real question.
Flops or floating point operations per second was used for measuring how fast you can do actual arithmetic.
(for non integers)
Hertz is the actual clock speed, and for integers operations you nowadays have an integer artithmetic unit that can do most operations in a single clock cycle. At least add and subtract. Multiplication is a little more expensive but not much, and division can be somewhat more expensive.
But real programs dont really limit themselves to integer math but need floating point numbers.
Floating point is the name of a different representation for numbers that are closer to describe rational numbers instead of integers.
There used to be a whole lot of different representations for these "rational" numbers but in the end ieee won out and everything today uses :
A Computer Science portal for geeks. It contains well written, well thought and well explained computer science and programming articles, quizzes and practice/competitive programming/company interview Questions.
www.geeksforgeeks.org
As you see, the representation is a LOT more complex than the integer case where you could just use a simple adder-circuit to add two numbers together.
In this case before you can even add the numbers you first need to normalize them so that they use the same exponent. etc.
Even a "simple" thing like addition becomes very complex and will require very many clockcycles to complete.
That is not even talking about things like multiplication, exponents, logarithms, trigonometric functions.
Some of the more complex operations on rational numbers/floating point numbers could take thousands or tens of thousands of clock cycles. Even with dedicated hardware.
Since the span in clock-cycles required between different operations spans across several/many orders of magnitude between
cheap and expensive operations hertz is not a practical measure of floating point performance. Henze using flops instead which is mostly kind of an average, middle of the road, of the more common operations. I think flops map closely to the cost of a multiplication, which is more expensive than an addition but less expensive than a division.
But even then, flops is too in-exact for some fields. Linear algebra, matrix multiplication, vector analysis, etc is all about a very specific operation where the atom is basically computing result = a * X + Y over and over and over.
Hence those folks measure performance in S/DAXPY. Single / Double precicion AX Plus Y.
On top of this there is an entire field halfway between mathematics and computer science called Numerical Analysis that is all about how to do floating point computations.