Microsoft Word Kurzweil, Ray The Singularity Is Near doc

Accelerating the Availability of Human-Level Personal Computing. 
Personal computers today provide more than 
10
9
cps. According to the projections in the "Exponential Growth of Computing" chart (p. 70), we will achieve 10
16
cps 
by 2025. However, there are several ways this timeline can be accelerated. Rather than using general-purpose 
processors, one can use application-specific integrated circuits (ASICs) to provide greater price-performance for very 
repetitive calculations. Such circuits already provide extremely high computational throughput for the repetitive 
calculations used in generating moving images in video games. ASICs can increase price-performance a thousandfold, 
cutting about eight years off the 2025 date. The varied programs that a simulation of the human brain will comprise 
will also include a great deal of repetition and thus will be amenable to ASIC implementation. The cerebellum, for 
example, repeats a basic wiring pattern billions of times. 
We will also be able to amplify the power of personal computers by harvesting the unused computation power of 
devices on the Internet. New communication paradigms such as "mesh" computing contemplate treating every device 
in the network as a node rather than just a "spoke."
43
In other words, instead of devices (such as personal computers 
and PDAs) merely sending information to and from nodes, each device will act as a node itself, sending information to 
and receiving information from every other device. That will create very robust, self-organizing communication 
networks. It will also make it easier for computers and other devices to tap unused CPU cycles of the devices in their 
region of the mesh. 
Currently at least 99 percent, if not 99.9 percent, of the computational capacity of all the computers on the Internet 
lies unused. Effectively harnessing this computation can provide another factor of 10
2
or 10
3
in increased price-
performance. For these reasons, it is reasonable to expect human brain capacity, at least in terms of hardware 
computational capacity, for one thousand dollars by around 2020. 

---

Yet another approach to accelerate the availability of human-level computation in a personal computer is to use 
transistors in their native "analog" mode. Many of the processes in the human brain are analog, not digital. Although 
we can emulate analog processes to any desired degree of accuracy with digital computation, we lose several orders of 
magnitude of efficiency in doing so. A single transistor can multiply two values represented as analog levels; doing so 
with digital circuits requires thousands of transistors. California Institute of Technology's Carver Mead has been 
pioneering this concept.
44
One disadvantage of Mead's approach is that the engineering design time required for such 
native analog computing is lengthy, so most researchers developing software to emulate regions of the brain usually 
prefer the rapid turnaround of software simulations.