Page 359 - System on Package_ Miniaturization of the Entire System

Integrated Chip-to-Chip Optoelectr onic SOP 331

In the remainder of this chapter we will consider the development of digital-optical
SOP technology as it applies to high-performance computing, servers, and routers.

6.4 Advantages of Optoelectronic SOP

6.4.1 Comparison of High-Speed Electrical and Optical Wiring Performance
There are three forces that drive SOP technology for computing, namely, (1) performance,
(2) power, and (3) cost. These translate into seven technology improvements that are
made possible by introducing optoelectronics architecture into digital systems. These
improvements are (1) nearly invariant (bandwidth × distance) product over the entire
network, (2) interconnect density that is far greater than copper densities, (3) negligible
crosstalk and insensitivity to simultaneous switching noise (SSN), (4) three-dimensional
optical wiring, (5) direct, high-speed processor-to-processor optical links that lead to
architectural simplicity by reducing the use of much slower copper bus lines, number
of printed circuit board (PCB) layers and vias, and capacitors, (6) greatly reduced node
crossing delays in multiprocessor networks by direct optical wiring and long-reach off-
chip synchronization, and (7) minimum number of components used for noise
suppression [15].
It is well known that the bandwidth of copper interconnects is intrinsically limited
by numerous factors such as skin effect, inductance, capacitance, and EM radiation, as
well as extrinsic factors such as the dielectric susceptibility of the insulator which may
cause frequency dispersion and signal attenuation, crosstalk, and power supply noise.
The intrinsic limitation of the (bandwidth × distance) product for unequalized copper
lines is summarized in Equation (6.1), where B max is the maximum bandwidth capacity,
A is the cross-sectional area of the copper line, and l is its length [16].

B ≤ . 24 x10 7 A/ρ Gb/s (6.1)
max
A graphical representation of Equation (6.1), namely the dependence of the bandwidth of
two unequalized copper transmission lines on cross-sectional area and distance, is shown
in Figure 6.4 and compared with the bandwidth for a multimode (MM) polymer waveguide
carrying a single wavelength (1310 nm) with an attenuation coefficient of –0.2 dB/cm.
In comparison, a single, long-distance optical fiber carrying 100 colors, each having
a 10-Gb/s capacity, has an aggregate bandwidth of 1 Tb/s. Under development are
100-Gb/s laser modulators and detectors with similar speed to support the Ethernet
infrastructure [2]. The ability to move massive quantities of data at several Tb/s over
long distances makes it possible to contemplate real-time distributed, task-specific,
parallel computing. Realizing the potential advantages of an optical network architecture
that links multiple nodes (processors, memory banks, I/O buffers), a number of authors
have simulated a variety of optical bus designs and data passing protocols in symmetric
multiprocessing (SMP) and massively parallel processing (MPP) machines. In SMP
(servers), all processors share the same memory via one or more buses, and each CPU
takes on the next available task. In MPP (supercomputers), problem segments are solved
in locked step. Each CPU contains its own copy of the operating system and application
and has access to its own memory. Tasks are assigned to each CPU, and communication
between MPP subsystems takes place via high-speed interconnects where the optics
solution can play a role.