He claimed anywhere from 3,300x in linear programming from 1988 to 2004, to a 75,000x speedup in algorithmic improvements in CPLEX from 1991 to 2007.
One of the points he drives home in his tables is that the algorithmic improvement isn't "dwarfed" by processor/machine improvments. Rather they multiply eachother. A 3300x speedup on the algorithm side, combines with a 2000x speedup on the machine side, for a total speedup of 6.6 million x.
So what took >76 days to compute in ~1990 now takes <1 second. But using the same algorithms from 1990 on today's computers, those computations might take 30 minutes instead of the 1 second that we are accustomed to.
0: https://www.math.uni-bielefeld.de/documenta/vol-ismp/25_bixb... [page 114]