IEEE Circuits and Systems Magazine - Q3 2022 - 53

way people live and work and require organizations to
rethink their teams, processes, and technologies to stay
competitive. Today, organizations across all geographies
and industries can innovate, create more customer
value and differentiation and compete on an equal
playing field. This new reality demands that enterprises
embrace digital transformation and pivot quickly or fail.
AI (Artificial Intelligence) is a critical part of the digital
transformation journey for many organizations. Smart
cities, wearable health technologies, smart energy
grids, autonomous vehicles, manufacturing, and agriculture
are just some of the key markets being transformed
by AI. He pointed out that how technology, open source
communities and new ways of collaborating are driving
business innovations like AI. AI investments across every
industry are accelerating to develop differentiated
services and gain competitive advantages. The difference
between complex hardware and software for key
elements like security, data is diminishing. Many businesses
are aware of the benefits, but there are a number
of challenges delaying their implementation plans.
This is the natural home of open source, providing the
components and foundation, and creating a space for innovators
to come together and share their great ideas in
a way that's self-sustainable.
Next Dr. Evgeni Gousev (Senior Director at Qualcomm)
gave a wonderful overview of tinyML: enabling
ultra-low power machine learning at the very edge.
In his talk, he covered many aspects such as tinyML
fundamentals, its markets and values, and gave many
examples. He discussed recent developments from
Qualcomm and provided information about the tinyML
foundation, ecosystem, projects and events and educational
activities. Dr. Gousev further defined tinyML as
machine learning architectures, techniques, tools, and
approaches capable of performing on-device analytics
for variety of sensing modalities (vision, audio, motion,
chemicals, etc.) at mW the power range or below, targeting
predominantly battery-operated devices. The key
tinyML growth drivers include more efficient hardware,
energy efficient algorithms, more mature software infrastructure,
tools, diverse ecosystems, growing the
number of applications, corporate investment, VC investment,
and increased start-ups. It is predicted that
1B tinyML devices would be shipped in 2024 and would
approach 5.4B in 2026. The growth is in double digits.
He described many useful applications of tinyML, such
as, voice recognition, environmental sensors, predictive
maintenance, gesture control, and augmented reality.
The final talk on the first day was given by Dr. Venkat
Thanvantri (VP of Machine Learning R&D at Cadence). He
presented advances in AI/ML for chip design. The emergence
of machine learning (ML) has unlocked many new
THIRD QUARTER 2022
applications and transformed user experiences. Electronics
Design Automation (EDA) is an application area
that delivers value by providing automation and abstraction.
ML technology is having a similarly transformative
impact on EDA, accelerating execution of algorithms, improving
quality of results, and now significantly improving
the productivity of users. In particular, he detailed
the use of Reinforcement Learning to automate and optimize
results for the digital design and signoff flow.
On the second day of the symposium, Dr. Suk Hwan
Lim (Executive VP at Samsung) addressed fine grained
domain specific architectures for diverse workloads.
He covered deep learning applications and workloads,
domain specific neural processing units/accelerators
(NPU), and future directions for NPUs. There are several
deep learning (DL) applications - speech recognition,
voice activation, text to speech, authentication, image
classification, object detection, semantic segmentation
and image processing, among others. Training is typically
implemented in the cloud, while inference is deployed
on the edge. The compute complexity varies by
5-7 orders of magnitude for these applications. Thus, he
categorized fine grained diverse processors for diverse
applications-micro NPU for audio/always-on applications,
general NPUs for small spatial resolutions with
deep network, and image processing NPUs for large
special resolutions with shallow networks. Significant
improvements in energy, area, efficiency and utilization
are needed and can be achieved through algorithms/
compiler/architecture/circuits.
Next, Prof. Song Han (Assistant Professor at MIT)
described the role of tinyML and how greener AI can
be achieved. TinyML and efficient deep learning make
AI greener and easily deployable to IoT. AI applications
can generate high power and have detrimental affects
on the environment. There is a push for data compression,
pruning, and other techniques to reduce computation
and thereby power. New models can be developed
to improve latency and accuracy. Manual design
is challenging and automation is needed. Prof. Han has
proposed and developed a hardware-aware neural network
search called " Once-for-All " . While computationally
expensive, it requires training only once and then
produces multiple models which can be used for inference.
This approach reduces the data usage and results
in less computation and hence a lower carbon footprint.
Also, sparse attention and progressive quantization
ideas are used to prune the tokens in Natural Language
Processing (NLP). Activation is the main bottleneck and
not the trainable parameter space. Activation minimization
leads to significant memory reduction in the IoT.
All the techniques described in his talk can help AI to
be greener.
IEEE CIRCUITS AND SYSTEMS MAGAZINE
53
IEEE Circuits and Systems Magazine - Q3 2022

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q3 2022
