A happy ending to my last rant:
I did some hacking around (with suggestions from the inimitable pabw), and started my kernel computation again. Last time, it ran for 9 days and failed when the power died. The ETA was around 38 days. This time.... 43 hours. Who now has the complete kernel matrix for 17 gigs of fMRI data? *Incrementally cached*? That'd be me. Bow, all ye data-matrix less peons.