Page 19 - A Practical Guide from Design Planning to Manufacturing

Preface xvii

The first of these questions has remained the biggest question facing
processor designers since the beginning of the industry. By the end of
2006, that is, when this book gets published, the highest transistor
count processors on the market should include more than 1 billion
devices. If the current rate of increase continues, a 10-billion device
processor is likely before 2015 and a 100-billion device processor by
2025. What will these processors be like? The most recent answer for
how to make use of more transistors is to put multiple processor cores
onto a single die. Does this mean that a 10-billion transistor processor
will merely be a combination of ten 1-billion transistor processors? This
is certainly possible, but a 100-billion transistor processor will almost
certainly not be a hundred core processor. At least today, most software
problems cannot be divided into this many separate pieces. Perhaps
new methods will be found, but it is likely that the number of cores in
future processors will be limited more by software than by hardware.
If most software applications will only be able to make use of a very
small number of cores, will each single core contain tens of billions of tran-
sistors? Design tools and methods of today are not up to creating the
design for such a processor. We may be moving from a time when proces-
sors designs are no longer determined by the limits of fabrication, but
instead by the limits of the design process itself. Perhaps the processor
will absorb the functionality of other computer components, as has hap-
pened in the past. A microprocessor with several general-purpose cores
as well as a graphics processor, memory controller, and even main
memory itself, all built into a single die, could make use of a very large
number of transistors indeed. Maybe this type of true “system-on-a-chip”
will be the future. In the past, it has always been feared that the end of
fabrication improvements was just a few years away. It is physically
impossible for the shrinking of transistors to continue at its current pace
forever, but every prediction so far of the end of scaling has been wrong.
Today, the problems of new hardware and software design methodologies
threaten to slow processor improvements before manufacturing limits.
The second critical question of power efficiency has received serious
attention from the industry only recently. Early “low-power” processors
were simply desktop designs operated at lower voltages and frequencies
in order to save power. Only recently has the rapidly growing popular-
ity of portable computing products led to the creation of a number of
processor designs intended from conception as low power. Power effi-
ciency has become even more important as high-performance desktops
and server processors have reached the limits of cost-effective power
delivery and cooling. Suddenly, 100-W server processors and 1-W embed-
ded processors have started to be designed for low power.
The industry has used a number of circuit designs and some fabrication
techniques to reduce power, but we have barely scratched the surface of