1 of 20

Developing a Neuromorphic Hub for USA�Ecosystem for Research, Infrastructure and Collaboration

NUAI LAB

March 11, 2025

NSF PI Meeting, Raleigh

PI:	Dhireesha Kudithipudi, PhD
Co-PIs:	Catherine Schuman, PhD�Gert Cauwenberghs, PhD
SP:	Vijay Janapa Reddi, PhD

2 of 20

To create next-generation AI systems capable of performing natural tasks with high energy efficiency

What we are curious about?

3 of 20

Why Neuro-Inspired Learning ?

Brains are working exemplar models

Efficient for information processing unstructured data

Hardware is not hardwired: Plasticity

LIFELONG/CONTINUAL LEARNING

Faults are indigenous

PARALLELISM, REDUNDANCY APPEAR TO COMPENSATE

Learning centric processing

WORKS WELL WITH IMBALANCED DATA SAMPLES

HUMAN BRAIN

4 of 20

The Cost of Computation in AI Hardware

Patterson, David, et al. "Carbon emissions and large neural network training." arXiv 2021.

AI compute is doubling

~3.5 months

Compute cost decreasing by 1 order

~4-12 years

Training cost of top models

~$10M

OpenAI Five (DOTA 2 video game)

Training equivalent ~ 45,000 years

5 of 20

The Cost of Computation in Brains

Levy WB, Calvert VG. Communication consumes 35 times more energy than computation in the human cortex, but both costs are needed to predict synapse number. Proceedings of the National Academy of Sciences. 2021 May 4;118(18):e2008173118.

~20W of Total Power

6 of 20

Neuromorphic Computing

Systems that closely emulate or mimic the structure and computational principles observed in the nervous system, through hardware and algorithm co-design.

NEUROSCIENCE

MACHINE LEARNING

COMPUTER ARCHITECTURES

NEUROMORPHIC ACCELERATORS

Energy Efficiency is the key driver for Neuromorphic Computing Systems

7 of 20

Neuromorphic Computing on Innovation Trigger

Neuromorphic systems have grown in scale from

a single neuron in the early

1960s

to many billions by

2025

Kudithipudi, Dhireesha, et al. "Neuromorphic computing at scale." Nature 637.8047 (2025): 801-812.

8 of 20

scale as the capacity of a system (inclusive of algorithms, hardware, architecture and infrastructure) to operate at the size, speed and energy required to address complex, real-world tasks.

Scale offers a measure to understand the density of the underlying hardware (with the well-known trajectory of Moore’s law) and the performance metric of supercomputers measuring floating-point operations per second (known as FLOPS).

Small Yet Mighty

Kudithipudi, Dhireesha, et al. "Neuromorphic computing at scale." Nature 637.8047 (2025): 801-812.

Why scale is important?

9 of 20

Features of Neuromorphic Computing

NEUROMORPHIC

LATE-STAGE MATURATION

FEATURE MATURATION

OF NEUROMORPHIC COMPUTING AT SCALE

NEUROSCIENTIFIC

NEUROMORPHIC

EARLY-STAGE MATURATION

RECONFIGURABLE

RECONFIGURATION

NEURAL PLASTICITY

Reconfiguring the routing and configuration of the neural processing elements.

SCALABLE

NEURAL SCALABILITY

NEURONAL ELEMENTS

Scaling the algorithm across multiple neuronal processing elements.

SPARSE

ACTIVITY

SPARSE CONNECTIVITY

Sparseness in the activity of the networks or in the connectivity between neurons.

RESOURCE AWARE

MEMORY

POWER

COMPUTE

Computing algorithms that are aware of and factor-in the memory, power and compute constraints.

HETEROGENEOUS

NEUROMORPHIC BOARDS

VARIOUS SENSORS

NERVOUS SYSTEM

Support for integrating a variety of computing platforms, sensors and actuators

ASYNCHRONOUS

NEURAL SPIKES

Asynchronous communications between neural processing elements.

DISTRIBUTED & HIERARCHICAL

CORTICAL REGIONS

Hierarchies of Computing Infrastructure.

10 of 20

What makes a neuromorphic system successful ?

"What we have before us are some breathtaking opportunities disguised as insoluble problems."

- John Gardner, 1965

11 of 20

THOR : The Neuromorphic Commons

AIMS

To develop and deploy THOR, a large-scale neuromorphic computing research infrastructure.

Foster collaborative research on biological intelligence, from perception to learning, enabling diverse applications.

The neuromorphic commons (THOR), is an open-source large-scale national neuromorphic computing resource on two distinct commercial-scale platforms

Open-Source Software Design Frameworks and Tools

Common benchmarks and community-driven competitions

Access to large-scale neuromorphic systems

HW/SW Codesign

and Development

12 of 20

Incentives

ROBOTICS

R&D Multiplier

Proof-of-Concepts

Neuromorphic Computing

Development Community

Mainstream Usage

Early & Easy Access To Technology

ACADEMIA

SMART CITIES

GREEN DATA

CENTERS

Community driven development of neuromorphic systems

Stages of Development

13 of 20

Neuromorphic Ecosystem

HW/SW

CO-DESIGN

MEMRISTOR CROSSBAR

RRAM DEVICE

ACCELERATOR ARRAY

PROCESSING ELEMENT

SOFTWARE CONFIGURATION

DIGITAL

ANALOG

LOW-LEVEL

Hardware Specific

ex. Assembly

MID-LEVEL

Core Performant Software

ex. CUDA, C

HIGH-LEVEL

Accessible,

Portable and Cross-Platform

ex. Lava, Fugu

HARDWARE

SOFTWARE

DEV BOARDS

e.g., Loihi, HICANN

NEUROMORPHIC CHIP

ALGORITHM LOGIC

NEUROMORPHIC ACCELERATOR

e.g., Genesis Chip

14 of 20

Early Demonstrations of Success

15 of 20

Team & Lead Organizations

COLLABORATING INSTITUTIONS

16 of 20

Research Domains

$4M

Physics

Lifesciences

AI / ML

Neuroscience

RESEARCH DOMAINS

First-of-its-kind, THOR infrastructure, enables richer understanding of computational models, AI/ML algorithms, neuroscience and neuromorphic hardware.

Healthcare

Our neuromorphic research explores the fundamental laws of nature, leverages biological insights, develops advanced learning algorithms, emulates brain function, and addresses critical health challenges.

17 of 20

How to evaluate research to neuromorphic domain ?

NeuroBench

A benchmark framework for neuromorphic algorithms and systems, which is collaboratively designed from an open community of researchers across industry and academia.

NeuroBench specifies a common set of tools and systematic methodology for inclusive benchmark measurement for both algorithms and hardware systems.

Yik, Jason, et al. "The neurobench framework for benchmarking neuromorphic computing algorithms and systems." Nature Communications 16.1 (2025): 1545.

19 of 20

Neuromorphic Community

Researchers and collaborators of the NUAI Lab

Dr. Andrey Kanaev (NSF)

Dr. Deepankar Medhi (NSF)

Acknowledgements