Hardware Reference
In-Depth Information
Chapter 1
Introduction to Network-on-Chip Design
Computing technology affects every aspect of our modern society and is a
major catalyst for innovation across different sectors. Semiconductor technology
and computer architecture has provided the necessary infrastructure on top of
which every computer system has been developed offering high performance
for computationally-intensive applications and low-energy operation for less
demanding ones. Innovation in the semiconductors industry provided more
transistors for roughly constant cost per chip, while computer architecture exploited
the available transistor budget and discovered innovative techniques to scale
systems' performance.
We have reached a point where transistor integration capacity will continue to
scale, though with limited performance and power benefit. Computer architects
reacted to this challenge with multicore architectures. The first systems devel-
oped followed an homogeneous architecture, while recent ones move gradually
to heterogeneous architectures that look like complex platform Systems-on-Chip
(SoCs) integrating in the same chip latency-optimized cores, throughput optimized
cores (like GPUs) and some specialized cores that together with the associated
memory hierarchies and memory controllers (mostly for off-chip DRAM) allows
them to cover the needs of many application domains. SoCs for mobile devices
were heterogeneous from the beginning including various specialized components
such as display controllers, camera interfaces, sensors, connectivity modules such
as Bluetooth, WiFi, FM radio, GNSS (Global Navigation Satellite System), and
multimedia subsystems. Programming such heterogeneous systems in a unified
manner is still an open challenge. Nevertheless, any revolutionary development in
heterogeneous systems programming should rely on a solid computation and com-
munication infrastructure that will aid and not limit the system-wide improvements.
Scalable interconnect architectures form the solid base on top of which
heterogeneous computing platforms and their unifying programming environments
will be developed; parallelism is all about cooperation that cannot be achieved
without the equivalent concurrency in communication. The interconnect implements
