Page 350 - System on Package_ Miniaturization of the Entire System

324 Cha pte r S i x

and routers will routinely process and share data at aggregate data rates of several
terabits per second using short-reach optical transceivers that employ optical
interconnects between chips on a multiprocessor board, and between boards “inside
the box.” The term “inside the box” generally refers to the inner portion of one of many
node enclosures that constitutes a supercomputer. It may also signify the enclosure that
defines the physical space of a stand-alone computer, such as a blade center server.
Optical interconnections “outside the box” will continue to use parallel, telecom-type
optical transceivers modified to accommodate high channel density.
Optoelectronic SOP provides options not available in conductive data transport:
(1) Since the optical signal bandwidth is independent of distance over distances of
several meters, extended memory may be placed in a separate box. (2) Optical links do
not require shielding and separation to reduce crosstalk to the extent that electrical lines
do. Less than 10 μm of cladding separation is sufficient at all data bandwidths. (3) Because
parallel optical links can be high density, potentially on a 60-μm pitch, and flexible,
typically capable of a 1-cm turn radius, they can, in principle, support direct processor-to-
processor wiring on a multiprocessor core module. (4) High-density flexible optical
interconnects will diminish the dependence of high-speed signal routing through the
backplane.
Optical-digital signaling “inside the box” will eventually migrate to optical-digital
signaling at the chip level. This paradigm shift will probably occur when the internal
processor clock speed increases to ≥10 GHz and clock synchronization and processor
temperature becomes difficult to manage. At that time, optimal codesign will integrate
on- and off-chip optical signaling at bit rates that are commensurate with that of an
internal processor clock. In this scenario, synchronization on- and off-chip will become
equally demanding. Alternatively, even if asynchronous processing is adopted, the
need for increasing bandwidth capacity will not diminish. One needs to be aware that,
in the near future, the total information transfer per year of the Internet is expected to
18
21
surpass the zettabyte (10 bytes) level. It is currently at the exabyte (10 bytes) level [1].
To meet this jump in information density, the Ethernet electrical and optical infrastructure
companies are busy developing 100-Gb/s capabilities worldwide [2].

6.2.2 RF-Optical Communication Systems
Optics is an effective enabling technology in the digital domain. Not only does optics
transcend the limitations of copper-based signaling (bandwidth × distance), but it also
transcends the limitation of impedance matching in three dimensions. Impedance matching
limits the transmission lines in two dimensions. Any long unshielded line along the signal
path acts as discontinuities, and impedance matching becomes difficult. Flexible, high-
speed, high-density copper interconnects increase the challenge because of radiative signal
losses, interference, and pulse distortion [3–4]. Flexible optical interconnects are subject to
fewer and less severe limitations: (1) The optical leakage is independent of bit rate. This is
not true for bent copper lines whose radiative losses increase with frequency [3–4]. For
example, polymer waveguides having a relatively small index contrast of 0.03 between
core and cladding can support a bend radius of 1 cm with about –0.12 dB loss [5]. (2) Straight
or curved copper lines are notorious for picking up and generating crosstalk primarily
through capacitive coupling. Design rules on minimum separation and minimum shielding
depending on the application are necessary. On the other hand, crosstalk between two
adjacent waveguides occurs when the evanescent field of one overlaps the other, which
translates to a distance of the order of a few wavelengths of light in the medium of propagation.