Yes, the 5-5-5 timings are a bit loose, but so are the 2.5-3-3 timings on DDR1. AMD does benefit from the lowest possible timings so running ddr2 667 at 3-2-2 will probably be better than running ddr2 800 at 4-4-4 (or 5-5-5).
I must say that socket AM2 is looking better and better as time passes. First we went from Anandtech's "horrible" performance numbers (which we never saw) to Tom's mediocre performance and finally we have HKEPC's slightly-better performance numbers. I'm not saying that it'll be a match to Conroe (which looks like it'll outclass the K8 for a while) but being within 10% in the mid-range does sound quite attainable. I agree that there is something obviously wrong with the memory controller because it's bandwidth seems to be only slightly better than a single-channel DDR2 800 setup would be. I'm not about to say that we should expect massive performance increases once this is corrected, but even a slight performance increase is better than none at all.
Also, according to the Inquirer (yes, I know...) AMD will begin mass production of 65nm parts in the next quarter to build inventories for a Q4 introduction (on the mobile/server side, I'd guess, so 65nm AM2 Opterons may hit the market this year).