Imagine using one of the world’s most powerful supercomputers, a machine that draws enormous amounts of power, to help design a far more efficient one. That is more or less Katie Schuman’s day job. Using systems such as Summit and Frontier hosted by the U.S. Department of Energy’s Oak Ridge National Laboratory, Schuman simulates neuromorphic architectures to understand how they might behave at scales that do not yet exist in hardware. The goal is to use today’s high-performance computing (HPC) systems to design a fundamentally different kind of computer, one that promises to operate with far lower energy demands.
Catherine “Katie” Schuman, Assistant Professor of Electrical Engineering and Computer Science, University of Tennessee, Knoxville.
Catherine “Katie” Schuman, Assistant Professor of Electrical Engineering and Computer Science, University of Tennessee, Knoxville
“I think it’s a fun use of the largest computers to say, ‘Maybe this isn’t the only way computing should be done,’” she says. “I do find it kind of ironic to be using these very large-scale systems that require a whole lot of power to design super-low-power computing for the future. But it’s kind of necessary. We have to.”
Past the irony, there is a practical question about the still-expanding field of neuromorphic computing: what does it take for a new computing architecture to move from promise to engineering practice? Schuman is determined to find that answer.
Catherine “Katie” Schuman is a faculty member in electrical engineering and computer science at the University of Tennessee, Knoxville, working at the intersection of neuromorphic computing, algorithms, software, and HPC-scale research. She co-authored the 2025 Nature review, “Neuromorphic Computing at Scale,” the work of 23 researchers across universities, national laboratories, and industry. She is also a co-investigator on THOR, the Neuromorphic Commons, an effort to broaden access to neuromorphic systems. Schuman’s efforts are about finding a practical path to a single question: how does neuromorphic computing move from a novel, interesting idea to something stitched into an actual, usable computing paradigm?
Partner Content
View All
Reliable Power-Path Design: Integrating MOSFETs, Diodes, TVS Devices, and Capacitors
Reliable Power-Path Design: Integrating MOSFETs, Diodes, TVS Devices, and Capacitors
By Unikeyic. 08.26.2026
From Days to Minutes: Accelerating 3D IC Debug with Agentic AI
From Days to Minutes: Accelerating 3D IC Debug with Agentic AI
By Zackary Glazewski, Founding AI Engineer, ChipAgents 08.26.2026
Why Automation Is Essential to Achieve EU CRA Compliance
Why Automation Is Essential to Achieve EU CRA Compliance
By Colin Duggan, CEO and co-founder, BG Networks 08.26.2026
Brain-influenced Computing
Schuman approaches the question with the requisite grounding: “I always say that neuromorphic is literally brain-shaped computing,” she says. “We’re taking inspiration from the brain, both in the literal architecture of the system and in how it operates.”
For engineers building next-generation neuromorphic architectures, this grounding is the foundation of the system itself. “It’s massively parallel. It has co-located processing and memory. It’s asynchronous and event-driven, and it can do very low-power operation across a variety of applications,” she says.
Schuman is clear that neuromorphic computing is not a replacement for today’s central or graphics processing units. Instead, neuromorphic offers another component in a mix that includes heterogeneous, workload-specific architectures. While today’s processors shuttle data between compute and memory under a global clock to execute instructions step-by-step in sequence, neuromorphic computing pairs the two. In the process, it sidesteps much of the power-hungry von Neumann bottleneck and allows many units of compute to act at once, each firing only when the right signal arrives.
So while neuromorphic computing is “brain-shaped,” its applicability extends well beyond neuroscience simulation. Schuman points to artificial intelligence, graph problems, optimization, and certain simulations as workloads that map well to the same massively parallel, event-driven model that defines neuromorphic hardware.
The Access Problem
Notably, that promise has outpaced the hardware. Schuman, who has worked in neuromorphic research for 15 years, has felt the lag, as she puts it, “basically my entire career.” For most of that time, neuromorphic systems have been few and hard to access, which has shifted much of the work to simulation. Modeling these architectures at any real scale is not something a laptop or a local cluster can do, so Schuman’s studies run on HPC leadership-class systems like Summit and Frontier. She turns to the largest machines available to explore hardware that does not yet broadly exist.
That reliance has a ceiling, and closing the gap has become a part of her mission.
“From the perspective of having to simulate these systems, to understand what they’re going to look like, to build the software and the algorithms for them, our existing computers aren’t built the same way,” she says. “That means it takes a whole lot longer to simulate something, and you need HPC-scale systems to understand how even a moderately sized neuromorphic system would behave. It’s not reasonable for me to be using Frontier and Summit every day to understand what the compute is going to look like.”
The point is not that she needs more time on the big machines, which Schuman contends has created “a dependency loop.” Instead, the field needs a better way to access the neuromorphic hardware itself.
“The hardware companies are waiting for there to be a killer application, but it’s really hard to understand how to build those applications without having hardware to prototype on,” she says. “You have to simulate what it’s going to look like, because our existing systems simply aren’t built the same.”
This maturity gap is wider than merely the chips because having the hardware alone does not make an architecture engineerable. Even where the devices exist, the surrounding support stack is currently thin. Whether the missing piece is programming abstractions, compilers, or workload-mapping tools, none of it becomes usable until engineers have the tools to do the job.
A Commons Modeled on HPC
Building this support stack is the purpose behind THOR, the Neuromorphic Commons. Schuman explained THOR aims to build a shared model that makes large-scale neuromorphic systems easier for researchers and students to access. THOR builds on guidance from the HPC community that preceded it and applies lessons learned from large-scale supercomputing practices today.
“We’re really trying to provide a user facility in much the same way these HPC systems are, so folks can access them,” Schuman says. “To date, this hasn’t been something an individual research lab at a university could acquire. In fact, even at national labs, relatively few DOE labs have acquired these sorts of neuromorphic systems.”
The bottom line is that shared systems create shared users, shared tooling, and shared ways of evaluating results, and, in time, a community forms around them.
“It also enables the community to come together to build larger-scale software and algorithm efforts, similar to what HPC has done in building the community around its systems,” she says.
The Field Needs Systems People
Deeper collaboration is critical to the success of neuromorphic computing. As Schuman puts it, “no one person has all the knowledge across the entire compute stack to build their own neuromorphic system, software and hardware, by themselves.” The 23-author Nature review is a clear signal that neuromorphic computing at scale requires broad input.
Schuman says that the practical conclusion is that more engineers need to get involved. The field has hardware and growing access to it, but it is short on the people working across the layers to make that hardware usable, especially at scale.
The TENNLab Neuromorphic Computing Research Group.
The TENNLab Neuromorphic Computing Research Group, which Schuman co-leads, at its all-hands meeting in June 2026 at the University of Tennessee, Knoxville. The group spans faculty, graduate students, and undergraduate researchers. Image Credit: TENNLab / University of Tennessee
“One of the big gaps we have is in programming languages, programming abstractions, and compilers,” she says. Few researchers work there, and that is precisely the expertise the electronics and computing communities hold in depth. “One of the things I’m always saying to folks in the HPC community is that we need their help.”
For systems engineers, this is familiar territory. Usable computing platforms are not made by silicon alone. New computing paradigms require smart and disciplined trade-offs.
This is the mission Schuman keeps returning to. Neuromorphic computing needs people who can reason across architectures, memory, parallelism, compilers, applications, and measurement. Many of them reside in the supercomputing and HPC communities. Both have spent decades building culture and shared practices around exactly those problems. The next stage will be for those leaders to find their way into neuromorphic and do for it what they did for the computing fields that came before.
“The HPC community is used to thinking about parallel computing, which is natively what’s happening on neuromorphic systems, so they’re already primed to think in the direction of what it would look like to use one of these novel compute platforms,” Schuman added. “I just want them to know that we’d like their help and would like to collaborate.”
Get Involved at SC26
SC26, the International Conference for High Performance Computing, Networking, Storage, and Analysis, takes place on November 15-20, 2026, at McCormick Place in Chicago. It is where HPC researchers, systems architects, and application developers gather to build community and common ground. The theme, “HPC Unites,” is an umbrella for the many communities advancing supercomputing in all its forms – the people working on AI and its infrastructure, quantum, networking, storage, and the novel architectures now entering the conversation, including programming models, compilers, and parallelism – expertise that neuromorphic computing needs. Whatever corner of computing you work in, if its future runs through large-scale systems, this is where those conversations happen. Registration is open, and the program is available here.