Math libraries are very often compared based on FLOPS. What information is being conveyed to me when I'm shown a plot of FLOPS vs size with sets of points for several different math libraries?
FLOPS as a measure of performance would make more sense to me if the comparison was between two implementations of the same algorithm or between the same software on two different pieces of hardware. I don't understand why it's an appropriate or popular way to compare things like matrix-matrix multiply.
Is the implication just that the underlying algorithms are nearly the same and the code that feeds the floating point units the fastest by minimizing overhead wins?
Examples abound.
http://eigen.tuxfamily.org/index.php?title=Benchmark
https://code.google.com/p/blaze-lib/wiki/Benchmarks
On the other hand, these LAPACK and Armadillo benchmarks use absolute time for a given operation, which makes more sense to me.
http://www.netlib.org/lapack/lug/node71.html
http://arma.sourceforge.net/speed.html
Relevant: