Sidebar
Main Menu
Home
Key Areas
Ageing
Asian Studies
Biomedical Science and Translational Medicine
Finance & Risk Management
Integrative Sustainability Solutions
Maritime
Materials Science
Smart Nation
Research Capabilities
Campus as a Living Lab
Shared Research Facilities
NUS Sustainable Futures
NUS Energy Solutions Hub (NESH)
Initiatives for Climate, Ocean & Nature (ICON)
Integrative Urban Solutions (IUS)
National Institute for Marine and Ocean Sciences (NIMOS)
Grant Call and Events
Research Office
Research Administration
Research Administration Resources
Integrated Research Information Management Suite (iRIMS)
iRIMS-infoReady
iRIMS-SponsoredProject (SP)
iRIMS - Fund Manager (FM)
iRIMS-IRB (Human Ethics)
iRIMS-IACUC (Animal Oversight)
iRIMS-IBC (Hazard Safety)
External Grant Management Systems
Integrated Grant Management System (IGMS)
iGrants
External Grants Indirect Research Cost (IRC) Recovery
Academic and Research Compliance and Integrity Office
Resources
HBRA E-Training
CITI RCR
Whistleblowing
Submit Tissue Registration
Amend Tissue Registration
Register Final Research Data
NHG ECOS Support
SingHealth ECOS Support
Institutional Review Board
NUS Institutional Animal Care and Use Committee (IACUC)
Research Data Management
Comparative Medicine
Research Concierge/Partnership
Research Agreements
Fellowships
NUS Fellows Programme (Southeast Asia)
Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship
Lee Kuan Yew Postdoctoral Fellowship
ODPRT - Jobs
Research Showcase
Research Features
Ignite Magazine
Researchers
News
Email
Library
Student
Staff
Home
Key Areas
Ageing
Asian Studies
Biomedical Science and Translational Medicine
Finance & Risk Management
Integrative Sustainability Solutions
Maritime
Materials Science
Smart Nation
Research Capabilities
Campus as a Living Lab
Shared Research Facilities
NUS Sustainable Futures
NUS Energy Solutions Hub (NESH)
Initiatives for Climate, Ocean & Nature (ICON)
Integrative Urban Solutions (IUS)
National Institute for Marine and Ocean Sciences (NIMOS)
Grant Call and Events
Research Office
Research Administration
Research Administration Resources
Integrated Research Information Management Suite (iRIMS)
iRIMS-infoReady
iRIMS-SponsoredProject (SP)
iRIMS - Fund Manager (FM)
iRIMS-IRB (Human Ethics)
iRIMS-IACUC (Animal Oversight)
iRIMS-IBC (Hazard Safety)
External Grant Management Systems
Integrated Grant Management System (IGMS)
iGrants
External Grants Indirect Research Cost (IRC) Recovery
Academic and Research Compliance and Integrity Office
Resources
HBRA E-Training
CITI RCR
Whistleblowing
Submit Tissue Registration
Amend Tissue Registration
Register Final Research Data
NHG ECOS Support
SingHealth ECOS Support
Institutional Review Board
NUS Institutional Animal Care and Use Committee (IACUC)
Research Data Management
Comparative Medicine
Research Concierge/Partnership
Research Agreements
Fellowships
NUS Fellows Programme (Southeast Asia)
Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship
Lee Kuan Yew Postdoctoral Fellowship
ODPRT - Jobs
Research Showcase
Research Features
Ignite Magazine
Researchers
Ignite Issue 03
11 Nov 2026
Every AI Chip Wastes Energy Shuttling Data. What if It Didn’t Have To?
Brain-inspired in-memory computing hardware development to reduce energy consumption and help power the next generation of AI
Professor Chen Wei
NUS Chemistry and Physics
Professor Ang Kah-Wee
NUS Electrical & Computer Engineering
Ignite Issue 03
11 Nov 2026
Every AI Chip Wastes Energy Shuttling Data. What if It Didn’t Have To?
Brain-inspired in-memory computing hardware development to reduce energy consumption and help power the next generation of AI
Professor Chen Wei
NUS Chemistry and Physics
Professor Ang Kah-Wee
NUS Electrical & Computer Engineering
The human brain, which runs on roughly the power of an LED light bulb, has inspired an emerging alternative to traditional chip designs, called in-memory computing. The brain’s synapses store and process information simultaneously, unlike modern computers, which shuttle data between distinct memory and processing units. In-memory computing cuts energy consumption by performing computations directly within the memory, eliminating the data shuttle.
However, this approach is a radical departure from what the semiconductor industry has relied on for over half a century, and it remains unproven at manufacturing scale. Professor Ang Kah-Wee from NUS Electrical and Computer Engineering and Professor Chen Wei from NUS Chemistry and Physics are tackling the problem from complementary angles: Prof Chen is investigating the material and device physics behind reliable two-dimensional (2D) material artificial synapses, while Prof Ang is developing wafer-scale integration strategies for scalable in-memory architectures.
The end of the data commute
“In the past, [the conventional computing] architecture served us very well... It’s only recently [that] this architecture failed to deliver,” noted Prof Ang. Data movement now accounts for a significant fraction of the energy consumed during AI inference, making memory bandwidth, rather than computation, the dominant bottleneck. Between 2018 and 2023, the size of frontier AI models grew roughly tenfold every year, with each leap demanding a corresponding surge in energy. With our current hardware, “we are finding it hard to keep up”, he said. In-memory computing promises dramatic energy savings – but how much depends on how it is implemented.
Digital in-memory methods can be more easily integrated with existing systems, but the efficiency gains are modest, since they still rely on binary processing: every piece of information must be represented as a long string of ones and zeros, each requiring its own transistor. Modern computers already have over a hundred billion energy-consuming transistors crammed onto a single microchip, Prof Ang commented.
Prof Chen and Prof Ang are both focused on analogue in-memory computing, which can be implemented with almost a thousand times fewer components, significantly improving energy efficiency. One promising device for this approach is the memristor, which changes its resistance based on the voltage applied to it – a phenomenon known as resistive switching. Just as synapses can adjust their strength in response to neuronal activity, a memristor’s conductance can be incrementally adjusted through successive voltage pulses, allowing it to store and update values.
In a conventional chip, the core mathematical operation behind neural networks, matrix-vector multiplication, takes thousands of sequential steps – each shuttling data between memory and processor. Memristors arranged in a grid, known as a crossbar array, can do it in just one: input voltages are applied, each memristor produces a current proportional to its stored value, and these currents are physically summed as they flow along each column; multiplication and addition are performed not by logic gates, but by the natural behaviour of electrical current.
Despite in-memory computing’s promise, however, the field is “still at the exploratory stage,” said Prof Ang, and there is no consensus yet on the best material for resistive switching. But this is not unusual, he noted. To replace silicon dioxide in transistor gates, the semiconductor industry spent roughly a decade evaluating an enormous number of candidate materials before converging on hafnium oxide.
Engineering artificial synapses from 2D materials
Already the target of heavy industry investment as a potential replacement for silicon in logic chips, ultra-thin 2D materials are showing strong potential as switching layers for in-memory computing.
In a study published in Nature Communications in October 2025, Prof Chen’s group demonstrated scalable arrays of molybdenum disulfide memtransistors that integrate transistor operation with memristive behaviour. By carefully engineering sulphur vacancies and tuning Schottky barriers – energy barriers at the metal–semiconductor interface – the team fine-tuned how easily current flows into each device, achieving a high switching ratio of up to 100,000 with stable analogue conductance updates, while only consuming approximately one trillionth of a joule per operation.
By emulating synaptic strength through adjustable conductance, the array achieved over 98 per cent accuracy in a neural network image-classification demonstration. Importantly, the devices maintained reliable performance even at extremely small scales – a promising sign for the dense, high-capacity arrays that real-world AI applications would demand.
Getting 2D materials industry-ready
However, complementary metal-oxide-semiconductor (CMOS) technology has been the cornerstone of chip fabrication for over sixty years, and commercial fabs are keen to squeeze their existing toolsets as far as they can before embracing something new, said Prof Ang. For in-memory computing to gain a foothold, researchers must show that their proposed materials can be grown reliably at wafer scale and transferred – without upending existing fabrication processes.
For Prof Chen, this means that 2D materials need uniform thickness and film quality, as even the tiniest variations can lead to large performance differences in analogue synaptic devices. Prof Ang also cautioned: “If your transfer [onto the target chip] is not done well, that layer can [have] wrinkles, bubbles, cracking.” Noteworthily, even microscopic defects can cause large variations in device behaviour, making it impossible to fabricate the millions of identical artificial synapses required for commercial AI chips.
To help overcome these manufacturing hurdles, Prof Chen’s group combines advanced material synthesis and characterisation techniques to better understand and control the properties of emerging 2D materials. Using molecular beam epitaxy (MBE) to grow these materials and scanning tunnelling microscopy (STM) to examine them at the atomic scale, the team identifies the factors that influence material quality, uniformity and reproducibility.
In a separate study published in Nature Communications in March 2025, Prof Ang’s team grew a three nanometres thick hafnium diselenide layer across an entire wafer and developed a metal-assisted dry transfer process to move it onto silicon sectors at low temperatures, preserving the film’s integrity and keeping the process compatible with existing manufacturing methods. Rather than demonstrating an isolated device, the work demonstrated a heterogeneous platform integrating over a thousand 2D memristors with silicon selector technology, representing one of the first practical pathways toward manufacturable in-memory hardware.
Rather than relying on software simulations, which can paper over real-world challenges like current leakage and signal noise, the team demonstrated the full system end-to-end in hardware – chip, circuit board, and control logic. It delivered 97.5 per cent accuracy on a pattern recognition task, just shy of the perfect score achievable by a conventional software model.
To further prove that in-memory architectures are foundry-ready, Prof Ang is now leading the Energy-Aware Accelerated Computing (EAC) programme. Running until 2030, the initiative brings together teams from across NUS and A*STAR’s Institute of Microelectronics to demonstrate new AI accelerator architectures at foundry scale that consume significantly lower energy than conventional designs.
However, to fully realise brain-inspired, in-memory hardware systems, the broader technology stack would have to be substantially redesigned, Prof Chen noted. The semiconductor industry is currently built entirely around digital architectures, and adapting fabrication processes for analogue, in-memory hardware will require coordinated investment across the industry.
But the need for change is undeniable. For AI to continue to advance, more energy-efficient hardware is compulsory, Prof Ang stressed. “If you cannot manage the power, you cannot scale,” he said. Just as silicon enabled the computing revolution of the past sixty years, energy-efficient in-memory hardware could underpin the next generation of AI systems.
This article was first published in
Ignite Magazine, Issue 03
.
More News
Home
Home
Chen Wei Ang Kah Wee Every AI Chip Wastes Energy Shuttling Data