The result of this research, announced this past summer, is
the 50-Gbps silicon photonics link. This link demonstrates all
of the components that are necessary to develop viable silicon
photonics products in the future. This includes not only
the silicon chips themselves, but packaging consistent with
high-volume manufacturing, assembly and connectorization.
In addition, we sought to demonstrate bandwidth scalability
with wavelength-division multiplexing, and to show a path to
low-cost, terabit-per-second optical links to meet the challenge
of the information explosion.
A wealth of data
Before describing the details of our 50-Gbps link, it is useful
to take a few moments to appreciate the phenomenal flood of
data at hand. This torrent of bits is the result of several technological developments, starting with the growth of the Internet
itself. Nearly 2 billion people use the Internet, a figure that
continues to grow rapidly. In 2014, the Internet will be four
times larger than it was in 2009. Internet usage models have
also expanded, enabling everything from cloud and cluster
networks to Web-enabled, data-hungry consumer applications
such as video on demand.
Companies have taken advantage of the Internet and its
capacity to collect, share and generate information. Businesses
from manufacturing and finance to health care and travel
now rely on massive quantities of data to fuel their operations.
Large retailers, for example, handle as many as a million customer transactions every hour, feeding databases estimated at
more than 2. 5 petabytes—167 times the storage of the books
in America’s Library of Congress. Furthermore, information
begets information. Databases such as these feed analytical
engines to deduce trends on customer buying habits or trends
in massive scientific datasets—a process that in turn produces
yet more data to share and store.
Consumers also contribute to the data explosion. Masses
of people are creating and sharing content on social networks
[ Low-cost optical interconnects ]
Courtesy of Intel Corp.
Optical inks
Optical
links
Shared
DRAM
bank
Board-level
multiprocessing
Low-cost optical interconnects will remove distance constraints for
high-bandwidth I/O, providing new options for system designers.
Integrated receiver chip
Integrated
transmitter
chip
Courtesy of Intel Corp.
Intel Labs’
50 Gbps silicon
photonics link.
as part of their daily routine. The most notable impact to
networks is with video. According to the Cisco Visual Networking Index, video already accounts for more than a third
of consumer Internet traffic. This will grow to 90 percent by
2014, with HD and 3-D video accounting for nearly half of
this traffic, and with applications such as video conferencing
and live streaming video rapidly growing as well.
Inevitably, this wave of data crashes upon the individual
computer systems that need to make sense of the bits and
bytes, systems whose computational demands are becoming
increasingly I/O limited. Multicore processors are packing
more computing capability into smaller spaces. Within a
server chassis, these high-performance processors are constrained both electrically and physically. As the demand for
higher data rates increases, traditional copper channels require
a closer proximity to processors, memory and I/O. This limits
both the number of memory channels and the number of
memory modules per channel, affecting bandwidth efficiency
and capacity per processor socket.
Rethinking systems designs
Optical interconnects will become a game changer if we can
combine the high data rates, reach and integration of multiple
photonic channels with prices that are practical for system-level interconnects. Eliminating the constraints of power and distance would provide the flexibility to revolutionize system-level
architecture. Servers would be able to expand memory beyond
what is physically and electrically feasible by connecting to
remote, optically attached memory systems without sacrificing
speed, thus maximizing both delivered bandwidth and capacity. Data centers would be able to maximize the performance
of multicore processors for such applications as large databases
or virtualized environments.
For manufacturers, this capability could lead to “
hyper-docks.” Imagine plugging an ultrathin laptop into an office
docking station that immediately gives it the performance of
a high-end work station through direct connections to additional memory and coprocessors. Or consider the future of
video displays. Today 1080p seems sufficient on most TVs—
but imagine a screen that fills your entire wall, something
you often see in science fiction stories, and using that wall
