The Alternative theme
is chartered to explore radically new models and modes of computation that have
the potential to simultaneously provide between 1-to-3 orders-of-magnitude improvement
in reliability and between 30%-to-75% improvement in energy-efficiency metrics
over traditional approaches, in the presence of extreme imperfections, e.g.,
raw error-rate of 20%, in nanoscale and post-silicon device and circuit
fabrics.
The Alternative theme
seeks to achieve these objectives by developing the foundations of computing in
the presence of statistical realities of the nanoscale era. We will explore
models of computation that achieve energy-efficiency and robustness by
consciously exploiting the statistical attributes inherent in emerging
nanoscale and post-silicon, device and circuit fabrics, i.e., statistical
behavior, and those of the emerging applications, i.e., statistical metrics of
performance. In particular, we seek to exploit the latter (statistical
application-level metrics) in order to accommodate the former (statistical
device/circuit behavior) in a systematic fashion. Our research derives its
inspiration from diverse areas such as statistical estimation and detection,
communications and information theory, and neuroscience. Our research targets emerging
applications in both commercial, such as sensor and mobile platform segments,
and defense sectors, such as target recognition, and ultra wideband data links.
The theme is
organized into two distinct research clusters: 1) stochastic computation, and
2) stochastic communication. Though distinct, each cluster adheres to the same
underlying philosophy of engineering the statistical behavior of the process,
device and circuit fabric to exploit the application-level performance metrics,
which also tend to be statistical in nature, such as quality of
behavior/service (QoS/QoB), signal-to-noise ratio (SNR), bit error-rate (BER)
and others. The Alternative theme will engage actively with the proposed MSD
devices center in developing and leveraging statistical behavioral models of
nanoscale circuit fabrics based on emerging carbon-based devices, in order to
study the effectiveness of its system-level techniques. The Stochastic
Computation cluster targets the design of reliable and energy-efficient
computational systems using highly unreliable nanoscale components by employing
statistical estimation and detection theory. These include the development of principles
of stochastic computation, their application to the design of stochastic
processors, and programming of such processors. The Stochastic Communication
cluster addresses the issue of designing reliable and energy-efficient
communication links and networks employing nanoscale circuit fabrics, including
both data transfer and synchronization. We will develop an
information-modulated view of communication links whereby robust and
energy-efficient data-transfer is achieved by treating all the processing
blocks in the transmitter to the receiver, particularly analog/mixed-signal, as
part of a composite unreliable channel. Modeling of the statistical behavior of
mixed-signal blocks for use in such links, and the use of on-off oscillator
arrays for low-energy synchronization will be explored.
