The first few tests are synthetic memory tests. They aren't a good indication of real-world application performance, but they can illustrate the impact of larger L2 cache on the memory subsystem.
The 600 series lands about where we'd expect overall, though it is a little bit slow in Sandra, possibly because of the extra overhead imposed by the larger L2 cache. In cachemem, which is not so aggressive about doing buffering and the like to get the absolute maximum throughput, the new 2MB L2 chips shine. The P4 does its own speculative prefetching of data into its L2 cache, and with a larger cache, it's able to achieve quite a bit more throughput.
Linpack's calculations are performed almost entirely in the L2 cache of the new chips. Notice how the Pentium 4 560's performance matches that of the 660 until we get to just under 1024K, and then it begins to drop off. The 660 just keeps on truckin'.
The larger cache doesn't do much for memory access latency, unfortunately.