The initial development of multicore processors owed much to the continued shrinking of CMOS lithography. It has long been known that increasing the size/width of a CPU core quickly reaches the realm of diminishing returns.8 So as the size of a state-of-the-art core shrinks to a small fraction of an economically viable die size, the options are:
- Produce a smaller chip
- Add lots of cache
- Add more cores
The option of adding more memory bandwidth typically adds significant cost outside of the processor chip, including revised motherboards (perhaps requiring more layers), additional DIMMs, etc. Because of these additional costs and the burden of socket incompatibility, the option of adding more memory bandwidth will be considered separately from options that involve only processor die size.
Figures 1 and 2 show some of the relative performance and performance/price metrics for these three options, assuming a 30% lithography shrink (i.e., 50% area shrink) which also allows a 17% frequency increase for the single core, but requires a 17% frequency decrease for the dual-core (to remain in the same power envelope). Note that the SPECfp_rate2000 benchmark consists of 14 individual tests that are modeled independently. These are summarized as a minimum speedup, median speedup, geometric mean speedup, and maximum speedup. For the dual-core processor option, both the single-core (uni) and dual-core (mp) speedups are estimated.
