I'd say that the algorithm matters more. As in, today's mathematicians know more about the subject.
I suspect the actual implementation of said algorithms probably achieves a lower % of peak performance than the older ones (though to be fair they are /much/ more complex algorithms).
>I suspect the actual implementation of said algorithms probably achieves a lower % of peak performance than the older ones
In my experience I have found the opposite to be the case. Most old maths libraries are written in FORTRAN (generally an order of magnitude or more slower than a comparable C/C++) and the implementations of standard algorithms are often sub-optimal and naive. I got the same impression when I compared arctangent (Taylor series) implementations in π programs and Minimax in chess (see Bernstein's program for the 704). I would guess that in the worst case they were 100x slower than what those machines could theoretically achieve. The C+inline asm libraries of today might be 2-10x slower at worst and some are even bottlenecked by memory. In this case I doubt the programs discussed in the 1989 paper, which were written in Prolog and FORTRAN, are exceptions to this.
IE: The software matters more. As in, today's programmers know more about this subject.