Innovative power recycling technique could revolutionize electronics.
As electronic devices become smaller and more powerful, managing their
energy consumption is crucial for extending battery life, reducing data
center operating costs and decreasing environmental impact.
Ioannis Savidis, PhD, associate professor of electrical and computer
engineering, and his team in the Integrated Circuits and Electronics (ICE)
Design and Analysis Lab are developing innovative techniques to recycle
power that would otherwise be wasted inside microprocessors and
systems-on-chips (SoCs).
Inside the microprocessors that power our devices, electricity flows even
when components are inactive. In the past, techniques like clock gating and
power gating – that is, powering down inactive components – have been used
to curb energy waste. “These methods save power, but leakage remains a major
unresolved issue,” noted Savidis.
Rather than simply shutting down idle circuits, Savidis’ technique
repurposes them to do useful work powering lower voltage components. This
allows new processing capabilities to be added at low cost.
“Recycling leakage for useful computation has the potential to significantly
move the needle on energy efficiency,” he explained.
Savidis’ technique works by dividing a microchip into active blocks that run
programs and idle blocks that leakage can be tapped from. Special switches
route current from idle blocks into a ‘virtual ground’ that powers a
low-energy subsystem, recycling energy rather than losing it as heat as
current is dumped to true ground.
“Imagine two buckets connected by a tube,” Savidis said. “When the top
bucket has spare water, it can flow to fill the bottom bucket.”
A key innovation enabling efficient leakage recycling is the use of an
algorithm that was developed by Savidis’ team called longest idle
time-leakage recycling (LIT-LR) that schedules which idle blocks to tap at
any time. The algorithm monitors downtime and prioritizes blocks based on
idleness. Blocks idle for long periods supply the leakage current, while
briefly inactive blocks are avoided to minimize unnecessary switching
between the recycling and non-recycling state, which is energy inefficient.
In addition, the needed transistor switches account for up to 12% of total
power in modern processors; therefore, improving switch efficiency unlocks
large additional energy savings.
“Intelligently pairing idle blocks with active components is crucial to
maximizing benefits,” explained Savidis. “Otherwise, you risk negatively
impacting performance as well as energy efficiency.”
In simulations, LIT-LR reduced switching energy by 25% compared to random
idle block selection. Peak power also fell by over 7% using the tailored
algorithm. Savidis’ team has powered circuits at one-third the normal
voltage using recycled leakage in simulations.
The enhanced efficiency unlocked by Savidis’ research provides wide-ranging
benefits for both manufacturers and end users of electronic devices.
For device makers, recycling leakage current allows adding more
computational capabilities without sacrificing battery life or generating
additional heat. Higher-performance laptops and smartphones can incorporate
more advanced features while maintaining portability and charge duration.
Leakage recycling also reduces costs for data center operators by allowing
computation at lower voltages for applications with low computing demand.
Consumers stand to gain extended battery life for portable electronics
before needing a recharge. Plus, power savings ultimately translate to lower
prices over time. “Greater energy efficiency benefits manufacturers
initially through expanded product capabilities, then consumers as those
gains scale up,” explained Savidis.
On the sustainability front, improved energy efficiency directly reduces the
carbon footprint of electronic devices by curbing wasted power usage.
“Improving efficiency allows us to do more with less,” noted Savidis, “which
benefits the environment long-term.”
Including these kinds of efficiencies in microprocessors and SoCs pushes the
boundaries of what has already become an increasingly complex field of
circuit design. Each new generation of chips must pack more computing into
smaller spaces while minimizing delays and improving power savings. To that
end, an area of emerging interest in Savidis’ research focuses on developing
machine learning techniques for circuit design.
Machine learning can provide more accurate early estimates of factors like
timing and interconnect impedance before a final design is completed. This
reduces the need for extensive late-stage simulations and design recycles,
where design engineers must return to earlier stages and modify the circuit
to correct for missed performance and power targets.
“With traditional design, you may only discover a timing violation in later
testing stages, and you have to go back and reiterate through your design to
be able to meet whatever timing constraints you might have,” Savidis
explained. This iterative process takes considerable time and resources.
To enable their research, Savidis’ team needs large datasets rep- resenting
circuits to train and validate models. They developed a flexible framework
that quickly generates circuit layouts using an automated physical design
process. By controlling parameters like component connectivity, they can
easily generate thousands of unique circuit graphs and process reports. This
data trains machine learning models models to predict timing, impedance,
power and other characteristics.
In a recent paper, Savidis showed neural networks trained on circuit graphs
can make performance predictions for a new unseen
chip design. Fine-tuning the networks with a small sample of data from the
new design improves accuracy. This highlights the value of their graph
framework for capturing topological features to develop machine learning
techniques.
The potential benefits of the practice are far-reaching. Even if machine
learning can’t fully automate tasks, the gains in efficiency can be
substantial. “The hope would be that machine learning might not deliver your
ultimate solution, as you still have to run certain checks that traditional
tools are required for,” Savidis emphasized. “But if you only have to go
through three iterations of a design instead of 15, that’s a significant
time savings, and you can get your product to market faster.”
Savidis’ innovative research on recycling wasted energy and applying machine
learning to circuit design exemplify his creative approach to
problem-solving. His cross-disciplinary team combines expertise in digital
and analog circuits, low-power design, machine learning and beyond to push
boundaries in efficiency and automation. While challenges remain, their
progress developing novel techniques like leakage recycling and machine
learning-enabled circuit design pave the way for more powerful and
sustainable electronics.
“By improving energy efficiency and leveraging AI, we can continue advancing
technology while ensuring it benefits both consumers and the planet,”
Savidis said.