While this is a great innovation, i can see it potentially causing more issues than it actually fixes...

if you've ever tried using an XL amazon instance to run a DB server (64GB high speed ram, but very low disk IO performance0 you end up getting into situations where most of your data is cached in the buffer , and as soon as you get a checkpoint the disk is sooooooooo slow that your checkpoints time out and you end up with T-Logs that grow out of control, especially when you have high concurency.

i'd love to see people tackling the "real bottleneck" as steve put it... lets get Disk I/O up to scratch.