An interesting article on HPCwire caught our attention this week, in which two Berkeley scientists argue that continued reliance on multicore technology as a means to further parallel processing schemes is due to hit the proverbial brick wall sooner or later. Instead, they tout a "manycore" approach that would place perhaps several thousand simple cores per socket, rather than the eight or so of current designs. As an example, they cite Nvidia's new CUDA GPU, which touts 128 general-purpose cores, or a recent proposal by Intel that would have 80 cores performing functions in the teraflop range.
AMD is coming out with its own advanced design, one that would place a pair of dual-core Opterons alongside two R600 Stream processors, one of the ATI GPUs that is likely to become a general-purpose processor should AMD continue development on the project.
All of this is happening at a time of tremendous change in the very make-up of what has been known as the microprocessor for the past 40 years. As one article in Electronic Design points out, the shift from polysilicon and silicon-dioxide gate and insulator construction to hafnium and Intel's "high-k" solution is nothing short of a comet hitting the world of chip design.
So the next time someone says to you that the limits of Moore's Law have been reached, tell them there's still quite a bit of room for silicon to breathe. And both short-term and long-term prospects look good.