We are presently observing a paradigm change in designing complex
SoC as it occurs roughly every twelve years due to the exponentially
increasing number of transistors on a chip. The present design discontinuity,
as all previous ones, is characterized by a move to a higher level
of abstraction. This is required to cope with the rapidly increasing design
costs. While the present paradigm change shares the move to a
higher level of abstraction with all previous ones, there exists also a key
difference.
For the first time advances in semiconductor manufacturing do not
lead to a corresponding increase in performance. At 65 nm and below
it is predicted that only a small portion of performance increase
will be attributed to shrinking geometries while the lion share is due to
innovative processor architectures. To substantiate this assertion it is
instructive to look at major drivers of the semiconductor industry: wireless
communications and multimedia. Both areas are characterized by
an exponentially increasing demand of computational power to process
the sophisticated algorithms necessary to optimally utilize the limited
resource bandwidth. The computational power cannot be provided in an
energy-efficient manner by traditional processor architectures, but only
by a massively parallel, heterogeneous architecture.
The promise of parallelism has fascinated researchers for a long time;
however, in the end the uniprocessor has prevailed. What is different this
time? In the past few years computing industry changed course when it
announced that its high performance processors would henceforth rely
on multiple cores. However, switching from sequential to modestly parallel
computing will make programming much more difficult without
rewarding this effort with dramatic improvements.
A valid question is: Why should massive parallel computing work
when modestly parallel computing is not the solution? The answer is:It will work only if one restricts the application of the multiprocessor to
a class of applications. In wireless communications the signal processing
task can be naturally partitioned and is (almost) periodic. The first
property allows to employ the powerful technique of task level parallel
processing on different computational elements. The second property
allows to temporally assign the task by an (almost) periodic scheduler,
thus avoiding the fundamental problems associated with multithreading.
The key building elements of the massively parallel SoC will be clusters
of application specific processors (ASIP) which make use of instructionlevel
parallelism, data-level parallelism and instruction fusion.
This book describes the automatic ASIP implementation from the architecture
description language LISA employing the tool suite ”Processor
Designer” of CoWare. The single most important feature of the
approach presented in this book is the efficient ASIP implementation
while preserving the full architectural design space at the same time.
This is achieved by introducing an intermediate representation between
the architectural description in LISA and the Register Transfer Level
commonly accepted as entry point for hardware implementation. The
LISA description allows to explicitly describing architectural properties
which can be exploited to perform powerful architectural optimizations.
The implementation efficiency has been demonstrated by numerous industrial
designs.
We hope that this book will be useful to the engineer and engineering
manager in industry who wants to learn about the implementation
efficiency of ASIPs by performing architectural optimizations. We also
hope that this book will be useful to academia actively engaged in this
fascinating research area.