Yes, file size is exactly what I was thinking of, as important. If one is working with a small 1k text file (lets say in DOS, to get Windows out of the picture), then the 8 vs. 16 MB cache isn't going to make much of a difference. If one is in a situation in which the data set could fit in a 16 MB cache, but not an 8 MB cache, then increasing the cache size will help increase cache hits.
Generally speaking, and without looking at all the specifics of data access, I would expect there to be a point where a database server could continue to make use of the bigger cache, where many home user applications might begin to see some diminishing returns wrt performance in creasing the size of the thing. dBase servers are also looking at a large number of dBase records (themselves a bunch of table entries related through a foreign key and whatever SQL statement was issued for the data retreival) with each record being much smaller. But with possibly thousands of people hitting an Oracle server, and it's disk array...
As to Tom's Hardware, I stopped going there for the most part when Tom Pabst had that whole spat with Brian Hooke when Hooke was still working at Id Software, and then the like 8 versions of appologies Pabst made, the first few not sounding very appologetic at all.
To make a long story short, Quake 3 was in beta at the time, and Tom used the timedemo from the beta to compare the TNT2 vs the Voodoo 3. Tom drew conclusions off the timedemo, at which point Hooke, as an Id Software programmer had said that the timedemo in the beta was broken, and doesn't produce reliable benchmark scores yet. It was also noticeable, as in some parts of the beta, the timedemo just sat in a corner looking at a bare wall (aka blank screen), and kept taking benchmark readings on it.
Tom pretty much dismissed all Hooke was saying, and came off as arrogant and deragatory, insisting that his benchmark and therefor his roundup of the cards was correct. This continued and the spat between Tom Pabst and Id Software escalated, and in the end it came out on the authority of John Carmack himself, that the beta code for the thing was busted and still needed more work, just as one of Carmack's fellow programmers had been trying to tell him, before Id Software would be willing to consider their own timedemo to be reliable for use. Up to that point, the timedemo wasn't considered the most important part of their software to bug fix, I would imagine. At this point, Tom Pabst was becomming increasingly appologetic, but was also trying to "save face" rather then just admit straight out that he was wrong.
It ended with Brian Hooke assuring the code got corrected, and then running the benchmarks himself, with Id Software releasing their own set of "official" benchmarks for the cards Tom Pabst was testing, based on the then corrected (to Id's programmers satisfaction), then still beta code. It also ended with me writting Tom Pabst off as a reliable source of information for the most part, and my having pretty much ceased the visits to his site, I used to make.