ShowDaily Day3 - Flipbook - Page 18
sponsored by
Diffusion
continued from page 17
performance in tasks such as defect
detection and support data-driven
decision-making. Unlabeled data
refers to raw, unprocessed data that
lacks meaningful tags, annotations,
or categories. Unlike labeled data,
which humans have explicitly categorized, unlabeled data requires
an algorithm to discover the hidden
patterns and structures on its own.
Examples of raw, unlabeled data
include any form of collected data,
such as images, videos, audio
recordings, or text documents, that
has not yet been given a specific
context.
GANs
Before we cover the use of AI in
chip design, let’s examine a category of AI known as Generative
AI or GenAI. This term refers to
algorithms and models created by
machine learning (ML), specifically
generative adversarial networks
(GANs), that are used to create
convincing images, audio, and video
of real people performing activities,
such as playing tennis.
In chip design, Gen AI can learns a
vast spectrum of Verilog codes and,
when given prompts with specifications, can produce these codes
automatically. This technology has
the potential to revolutionize many
industries by optimizing results
that couldn’t have been achieved
previously, as well as automating,
thus increasing user productivity
substantially.
While GANs produce a highly
realistic and diverse sample, a
different approach –Diffusion
Models - is used for fine-grained
control and a diversity of image
results. Diffusion models are still
generative, which are trained by
attempting to generate images as
18 | Thursday, October 9
close as possible to a given training
data set. The training process
involves adding Gaussian noise to
the data and learning how to recover
data through the denoising process.
This way, the model can correct
itself over these small steps, eventually producing better samples.
The aim of all machine learning-based GenAI technologies is
to explore more possibilities than
could be done efficiently by humans
in a traditional setting. It involves
training an algorithm to make
decisions based on previous data
or patterns, thereby enabling the
system to generate better and more
accurate outputs. This approach is
particularly convenient for tasks
with specific parameters or weights,
such as the design of semiconductor chips. GenAI algorithms can
augment users’ work with more and
better scenarios, ultimately leading
to improved results and better
products.
Importance of diffusion
models in chip design
Now, let’s return to the design
of chips, where GenAI can help
explore architectural possibilities,
including fine-tuning place and
route (P&R) settings to achieve
better power, performance, and area
(PPA).
The denoising process used in
Diffusion modeling has become
a popular approach in GenAI.
Using incremental, active learning,
Diffusion models increase efficiency by automating processes that
generally require extensive manual
effort, such as P&R activities in chip
and board designs or optimizing
workflows. Some have noticed a 10X
or more productivity improvement.
In addition to greater layout
efficiency, diffusion models also
yield better results. GenAI can help
humans generate novel ideas and
solutions by exploring a more comprehensive range of possibilities than
a human could reasonably consider
previously.
All of the above lead to better
decision-making by enabling engineers to make informed choices
based on data-driven insights and
predictions. To be fully realized,
these benefits must work with the
existing chip design and verification
processes, all built upon a foundation of internal and third-party
intellectual property (IP).
Not surprisingly, AI IP platforms
have been developed to allow SoC
developers to design and deliver
optimal solutions for a wide range
of applications and markets.
Examples of these platforms with
such IP options include those that
provide a significant computational
offload from any host processor
and can scale to hundreds of TOPS
(Trillions or Tera Operations per
Second) for a multicore system. AI
IP platforms are also used to target
energy-efficient on-device-to-edge
AI processing that is critical to
support (AI/ML) SoCs, including
intelligent sensors, IoT, embedded
vision, AR/VR, and ADAS.
However, these benefits come at a
cost. Success depends upon a comprehensive approach that involves
collecting and preprocessing data,
training machine learning models,
and integrating these models into
the design workflow. Simulation
and analysis tools are crucial for
validating and refining designs.
Furthermore, designers and layout
professionals must adapt to new
workflows that embrace a datadriven approach to design.
www.semiconductordigest.com
