Hi all.
I've been doing some research into AI, and have found some interesting stuff.
Seems that most if not all new CPUs are now so complex and optimised that hundreds of "rip up/retry" cycles are needed just to get the CPU to work at all. This is another reason why overclocking the new chips is rarely effective.
So you have a CPU that has been designed by a computer, checked over (briefly) by designers, sent back n times for "repairs" and then finally masked and built.
Obviously errors can and do creep in, the AMD "Phenom" quad-core series are on the verge of a product recall due to an unrepairable performance-sapping erratum (15%!) and are probably going to have to be scrapped en masse.
The main issue on fast CPUs is interference between adjacent lines on multiple layers, and overclocking just makes things worse. However if you don't mind the odd one-in-1B calculation being inaccurate then this is not an issue.
For AI use, randomness can be a good thing (see "New Scientist" article on random interference on neural nets generating "ideas")
I predict that sooner or later, one or more of the new CPUs will display emergent properties such as generating a coherent pattern from random "noise" fed into the chip. In fact such a test may well reveal ways to increase performance so it is probably already being done.
Perhaps AMD shouldn't scrap these chips just yet, sell them to the scientific community for research use :)
-A