As
devices become increasingly embedded within the physical infrastructure around
us, they interconnect wirelessly to form what industry has dubbed the Internet
of Things (IoT). Such IoTs
imbue intelligence into our physical world and have the potential to transform
the future of cities, realizing diverse applications in areas such automotive
electronics, smart home devices, and intelligent transportation systems.
IoT sensor nodes and gateways are faced
with intense power constraints, with some deployments relying on solar power or
have casings that make fan cooling infeasible. As a result, many of these
devices tend to just gather the sensor inputs and back-haul them through the
Internet or dedicated cables to back-end servers for processing. This affects
response time for real-time applications, as well as the sensing fidelity.
Real-world IoT deployments face the need for higher
performance sensor nodes and gateways that can compute on the sensory data
in-situ.
This
project pulls together expertise spanning real-world IoT
application development, computer-aided design, computer architecture, to
low-power circuits, with the aim of designing next-generation high-performance
sensor gateway multi-core SoCs. Our approach is to
customize and automate. We will aggressively customize each component of the SoC to substantially lower its power – each processors’
width, functional units, instruction set will be tailored to the specific
applications; all the datapaths communicating between
the various SoC blocks, such as its width, frequency,
connections, will be tuned to fit the applications’ characteristics, and the
entire chip will be run at near threshold voltage, matching the applications’
required sampling frequency. All this customization will be intractable without
automation. Hence, we propose a design automation toolchain that will take as
input our IoT driving application programs, search
the design space of customization options, and automatically generate the
resulting SoC chip design tailored for our IoT applications.