Moore’s law and the silicon revolution - The Computing Universe: A Journey through a Revolution - page 134

Fig. 7.18. ICs are produced by a compli-
cated process called photolithography.
The technique is similar to traditional
photography but in this case we use a
silicon wafer instead of a film of
emulsion.

SiO 2
Si

Forming an SiO 2 layer

Applying photoresist

UV Light

Mask

Development

Exposure

Photoresist removal

Etching

Morris Chang completely changed the landscape of the semiconductor
industry. He enabled start-ups to start with a few million dollars rather than a
few hundred million. That makes a huge difference. 14

Jen-Hsun Huang, a co-founder of Nvidia, credits TSMC with enabling all sorts
of creative ideas in areas such as networking, consumer electronics, com-
puters, and automotive technology to be turned into successful companies
because “the barriers to getting your chips built, to realizing your imagination,
disappeared.” 15

The end of the free lunch: parallel computing and
the multicore challenge

As Moore's law predicted, designers and manufacturers have delivered
smaller, faster chips requiring less power for more than four decades. But
we have now reached the length scale at which the transistor's gate insula-
tor is only a few atoms thick. Because the transistor is not a perfect switch,
it leaks some current even when it is in the turned off state. As the transistor
size decreases, if we continue to scale down the voltage, the current leakage
increases exponentially. To keep this leakage under control, the voltage can no
longer be scaled down with the dimensions of the chip. We can still shrink the
size of the transistors and place more of them on a chip, but they will not be
much faster than current generation transistors because the insulating silicon
dioxide layer cannot get thinner and the power consumption of the chips lim-
its our ability to clock the chip as fast as we could. As a result, chip architects
have developed multicore architectures with multiple CPUs integrated on a sin-
gle chip. Dual-core chips have been in widespread use for some time now, and
quad-core chips are increasingly common. Eight-core chips are now available
and the industry is experimenting with chips containing tens or even hundreds
of cores ( Fig. 7.19 ).
Performance improvement must now come from writing software that uses
multiple cores together to solve a problem. For many types of applications it is