The precursor to this technology has been around for some time. In fact, I'm currently designing something using it. Field Programmable Gate Arrays (FPGAs). It's a collection of unconnected logic blocks on a chip. You write a "program" that defines the functionality you want, and translation software figures out how to implement it in the logic available on the chip. You program the chip with the connections the software creates for you and, presto, you've got a custom chip. You can create anything you want, within the limitations of the number of logic blocks available and the input/output pins. Some people have even implemented replacements for obsolete microprocessors so that old software can continue to run. Some really fancy designs will even load different connections "on the fly" so that the FPGA can take on different roles at different times. Very cool stuff.
See post by stingerman
From Dec 2003 wonder why the delay in the announcement?
The cache supposedly has been doubled. As far as yield, IBM is using their new eFuse technology in the 90NM core allowing them to manufacture 970's with redundant critical circuitry, which will allow them to switch to a hot swap section of the processor, if there is a manufacturing flaw in a more vulnerable section. These electronic fuses are implemented to improve yields. At 90NM they have a lot more room for this kind of creative design.